Skip to content

GitLab

"...git@developer.sourcefind.cn:lacacy/qwen_lmdeploy.git" did not exist on "b239346701bd8d9cbc2ba1a2f5053cb1e1d671b5"
Commit 09117e54 authored by quyuan's avatar quyuan
Browse files

add ci

parent c8d7d97d
Hide whitespace changes
Inline Side-by-side
Showing with 1396 additions and 0 deletions
tests/benchmark/benchmark.py 0 → 100644
View file @ 09117e54
"""
bench
"""
import os
import shutil
import json
import calculate_score
code_path = os.environ.get('GITHUB_WORKSPACE')
#评测集存放路径
pdf_dev_path = "datasets/"
#magicpdf跑测结果
pdf_res_path = "/tmp/magic-pdf"
def test_cli():
"""
test pdf-command cli
"""
rm_cmd = f"rm -rf {pdf_res_path}"
os.system(rm_cmd)
os.makedirs(pdf_res_path)
cmd = f'magic-pdf pdf-command --pdf {os.path.join(pdf_dev_path, "mineru")} --inside_model true'
os.system(cmd)
for root, dirs, files in os.walk(pdf_res_path):
for magic_file in files:
target_dir = os.path.join(pdf_dev_path, "mineru")
if magic_file.endswith(".md"):
source_file = os.path.join(root, magic_file)
target_file = os.path.join(pdf_dev_path, "mineru", magic_file)
if not os.path.exists(target_dir):
os.makedirs(target_dir)
shutil.copy(source_file, target_file)
def get_score():
"""
get score
"""
data_path = os.path.join(pdf_dev_path, "ci")
score = calculate_score.Scoring(os.path.join(data_path, "result.json"))
score.calculate_similarity_total("mineru", data_path)
res = score.summary_scores()
return res
def ci_ben():
"""
ci benchmark
"""
try:
fr = open(os.path.join(pdf_dev_path, "result.json"), "r", encoding="utf-8")
lines = fr.readlines()
last_line = lines[-1].strip()
last_score = json.loads(last_line)
print ("last_score:", last_score)
last_simscore = last_score["average_sim_score"]
last_editdistance = last_score["average_edit_distance"]
last_bleu = last_score["average_bleu_score"]
except IOError:
print ("result.json not exist")
test_cli()
os.system(f"python pre_clean.py --tool_name mineru --download_dir {pdf_dev_path}")
now_score = get_score()
print ("now_score:", now_score)
now_simscore = now_score["average_sim_score"]
now_editdistance = now_score["average_edit_distance"]
now_bleu = now_score["average_bleu_score"]
assert last_simscore <= now_simscore
assert last_editdistance <= now_editdistance
assert last_bleu <= now_bleu
if __name__ == "__main__":
os.system("sh env.sh")
ci_ben()
tests/benchmark/calculate_score.py 0 → 100644
View file @ 09117e54
"""
calculate_score
"""
import os
import re
import json
import scoring
from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
from nltk.tokenize import word_tokenize
from Levenshtein import distance
class Scoring:
"""
calculate_score
"""
def __init__(self, result_path):
"""
init
"""
self.edit_distances = []
self.bleu_scores = []
self.sim_scores = []
self.filenames = []
self.score_dict = {}
self.anntion_cnt = 0
self.fw = open(result_path, "w+", encoding='utf-8')
def simple_bleu_score(self, candidate, reference):
"""
get bleu score
"""
candidate_tokens = word_tokenize(candidate)
reference_tokens = word_tokenize(reference)
return sentence_bleu([reference_tokens], candidate_tokens, smoothing_function=SmoothingFunction().method1)
def preprocess_string(self, s):
"""
preprocess_string
"""
sub_enter = re.sub(r'\n+', '\n', s)
return re.sub(r' ', ' ', sub_enter)
def calculate_similarity(self, annotion, actual, tool_type):
"""
calculate_similarity
"""
class_dict = {}
edit_distances = []
bleu_scores = []
sim_scores = list()
total_file = 0
for filename in os.listdir(annotion):
if filename.endswith('.md') and not filename.startswith('.'):
total_file = total_file + 1
with open(os.path.join(annotion, filename), 'r', encoding='utf-8') as file_a:
content_a = file_a.read()
self.anntion_cnt = self.anntion_cnt + 1
filepath_b = os.path.join(actual, filename)
if os.path.exists(filepath_b):
with open(filepath_b, 'r', encoding='utf-8') as file_b:
content_b = file_b.read()
self.filenames.append(filename)
edit_dist = distance(self.preprocess_string(content_b),self.preprocess_string(content_a)) / max(len(content_a), len(content_b))
self.edit_distances.append(edit_dist)
edit_distances.append(edit_dist)
bleu_score = self.simple_bleu_score(content_b, content_a)
bleu_scores.append(bleu_score)
self.bleu_scores.append(bleu_score)
score = scoring.score_text(content_b, content_a)
sim_scores.append(score)
self.sim_scores.append(score)
class_dict[filename] = {"edit_dist": edit_dist, "bleu_score": bleu_score, "sim_score": score}
self.score_dict[filename] = {"edit_dist": edit_dist, "bleu_score": bleu_score, "sim_score": score}
else:
print(f"File {filename} not found in actual directory.")
class_average_edit_distance = sum(edit_distances) / len(edit_distances) if edit_distances else 0
class_average_bleu_score = sum(bleu_scores) / len(bleu_scores) if bleu_scores else 0
class_average_sim_score = sum(sim_scores) / len(sim_scores) if sim_scores else 0
self.fw.write(json.dumps(class_dict, ensure_ascii=False) + "\n")
ratio = len(class_dict)/total_file
self.fw.write(f"{tool_type} extract ratio: {ratio}" + "\n")
self.fw.write(f"{tool_type} Average Levenshtein Distance: {class_average_edit_distance}" + "\n")
self.fw.write(f"{tool_type} Average BLEU Score: {class_average_bleu_score}" + "\n")
self.fw.write(f"{tool_type} Average Sim Score: {class_average_sim_score}" + "\n")
print (f"{tool_type} extract ratio: {ratio}")
print (f"{tool_type} Average Levenshtein Distance: {class_average_edit_distance}")
print (f"{tool_type} Average BLEU Score: {class_average_bleu_score}")
print (f"{tool_type} Average Sim Score: {class_average_sim_score}")
return self.score_dict
def summary_scores(self):
"""
calculate the average of edit distance, bleu score and sim score
"""
over_all_dict = dict()
average_edit_distance = sum(self.edit_distances) / len(self.edit_distances) if self.edit_distances else 0
average_bleu_score = sum(self.bleu_scores) / len(self.bleu_scores) if self.bleu_scores else 0
average_sim_score = sum(self.sim_scores) / len(self.sim_scores) if self.sim_scores else 0
over_all_dict["average_edit_distance"] = average_edit_distance
over_all_dict["average_bleu_score"] = average_bleu_score
over_all_dict["average_sim_score"] = average_sim_score
self.fw.write(json.dumps(over_all_dict, ensure_ascii=False) + "\n")
return over_all_dict
def calculate_similarity_total(self, tool_type, download_dir):
"""
calculate the average of edit distance, bleu score and sim score
"""
annotion = os.path.join(download_dir, "annotations", "cleaned")
actual = os.path.join(download_dir, tool_type, "cleaned")
score = self.calculate_similarity(annotion, actual, tool_type)
return score
tests/benchmark/datasets/annotations/cleaned/cleaned_academic_literature_0b2c9c91f5232541a7ace8984df306b2.md 0 → 100644
View file @ 09117e54
# Characterization of severely deformed new composites fabricated by powder metallurgy including a stage of mechanical alloying
H. Ashuri, A. Hassani*<br>Faculty of Materials Science and Engineering, Semnan University, Semnan 35131-19111, Iran
## ARTICLE INFO
## Article history:
Received 11 February 2014
Received in revised form 11 June 2014
Accepted 4 August 2014
Available online 12 August 2014
## Keywords:
Nanocomposite
Mechanical alloying
Twist extrusion
Powder metallurgy
#### Abstract
Mechanical properties of new composites having a binary matrix of $\mathrm{Al}-4 \mathrm{Cu}$ reinforced with $\mathrm{TiO}_{2}$ nano particles were investigated. The composites which consisted of $2 \mathrm{wt} \%$ and $8 \mathrm{wt} \%$ of $\mathrm{TiO}_{2}$ reinforcement particles, were fabricated using mechanical alloying and a powder metallurgy route. Morphology, phases and compounds formed during ball milling and densification of samples were studied. With increasing percentages of the reinforcement particles, mechanical properties of the composites were enhanced. Microstructural evolution and mechanical properties changes of the composites after application of twist extrusion (TE), as a severe plastic deformation (SPD) process, were also investigated. It was revealed that the more TE passes the higher hardness and yield strength obtained. In addition, increasing TE passes, led to occurrence of a more homogeneous distribution of the reinforcement particles within the structure, and development of an ultrafine-grained nano-structure. The maximum allowable number of TE passes was found to be four, above which the materials failed.
## 1. Introduction
In recent decades, aluminum matrix composites (AMC) with discontinuous reinforcements have vastly been attracted by different industries due to their good mechanical properties. Large number of manufacture routes have been developed to produce these materials among which powder metallurgy (PM) routes have been more considered with several causes. First, in powder metallurgy a controlled phase microstructure can be achieved. On the other hand, lower temperatures used in PM processes make the interphase kinetics be precisely controlled. In PM routes, the powders of elements and alloys are used which might be more inexpensive, and of course, much more effective in reinforcement of the composites. Traditional stages of PM-AMCs fabrication include mixing and blending the powders; degassing the solidified product in vacuum; homogenizing through hot pressing or hot isostatic pressing (HIP) [1].
AMCs are widely used in automotive, aerospace and transport industries because of their light weight, high elastic modulus, improved strength and good wear resistance. Strength and wear resistance of these materials are strongly dependent on volume fraction, size and type of reinforcement particles. They are well established that compared to their un-reinforced matrix alloys show higher wear resistance. AMCs with ceramic particles including $\mathrm{SiC}, \mathrm{TiC}, \mathrm{C}_{4} \mathrm{~B}, \mathrm{TiB}_{2}$ and $\mathrm{Al}_{2} \mathrm{O}_{3}$ are relatively easy to process and, in comparison with fiber-reinforced composites, are nearly isotropic [2].
Particulate AMCs have introduced most wide spread applications and hold the greatest promise for future growth because of their tailored properties, low cost-effectiveness and high volume production methods [3]. Aluminum matrix composites are known to be hard materials exhibiting a low forming capacity through the conventional techniques. Nevertheless, many promising attempts have been made to produce Al composites with a high potential of being formed plastically and even superplastically while their strength is retained [4].
Mechanical alloying (MA) is an interesting powder metallurgy route for producing of powders with high homogeneity and uniformity. This technique is very effective in dispersion of reinforcement particles and enhances grain refinement, which induces an increase of strength and hardness [5].
In recent years, manifestation of severe plastic deformation (SPD) methods in material science has shed light on new prospects in achieving a unique combination of high strength and ductility [6] as well as attaining ultrafine-grained materials with improved properties. SPD is a family of metal forming techniques that use extensive hydrostatic pressure to impose a very high strain on bulk solids, producing exceptional grain refinement without introducing any significant change in the overall dimensions of the sample[7]. Several different SPD techniques are now available; these include high-pressure torsion (HPT) [8], equal channel angular pressing (ECAP) [9], multi-directional forging (MDF) [10], accumulative roll-bonding (ARB) [11], repetitive corrugation and strengthening (RCS) [12], spread extrusion (SE) [13], simple shear extrusion (SSE) [14] and twist extrusion (TE) [15,16]. The SPD products have much higher structural efficiency in comparison with their coarsegrained counterparts. However, high cost-effectiveness of most SPD methods is a central drawback to produce such materials in high quantities. Therefore, development of new SPD methods to tackle cost problem is important.
In 1999, Beygelzimer proposed a severe plastic deformation process that became known as Twist Extrusion (TE) [17]. This process can change the structure of materials, significantly improving some of their physical and mechanical properties and, even in certain cases, gaining new properties. TE works by extruding a prism specimen through a matrix whose profile consists of two prismlike regions separated by a twist passage. The extruded material undergoes an intense shift, with the properties that the final cross-section of the specimen is identical to the initial cross-section [18]. These properties allow for a repeated extrusion that accumulates the value of deformation. TE is carried out under high hydrostatic pressure in the center of deformation which is created by applying anti-pressure (back pressure) to the specimen when it exits the matrix. It is possible to produce more isotropic and homogeneous deformation by turning the samples $90^{\circ}$ in each consecutive deformation or alternatively, make the use of consecutive clockwise-anticlockwise-clockwise twists [19]. A comparison between TE and the two most widely used SPD methods, ECAE and HPT, reveals that firstly, TE provides some advantages over ECAE such as the ability to extrude the hollow parts and the rectangular cross-sections [6]. Secondly, HPT involves order of magnitude higher pressures than in any other SPD process which provides attainment of uniquely high strains and formation of ultrafine grained structures. From another point of view, twist extrusion that combines extrusion with torsion, was introduced to tackle the insufficiency of HPT, that is, its being limited to laboratory conditions due to small size of the samples [19]. There are currently three main application areas of TE: (a) obtaining ultrafine grained crystalline and nano-crystalline structures in bulk specimens, (b) increasing the plasticity of secondary non-ferrous metals and alloys, which allows one to significantly broaden the range of production, (c) obtaining bulk specimens by consolidating porous materials which allows one to create substantially different, new compositions with unique characteristics [20].
In TE, strain distribution along the cross-section of the specimen is inhomogeneous; getting away from the axe, plastic strain increases, thus, the grains being finer. The microstructural inhomogeneity leads to inhomogeneities in the mechanical properties of the composite; the central area of the cross section having lowest strength. It is expected that with increasing the number of TE cycles, the microstructure becomes uniform [21].
In the present study, a powder metallurgy route combined with mechanical alloying was employed to produce some particulate $\mathrm{Al}-\mathrm{Cu} / \mathrm{TiO}_{2}$ composites with low $\mathrm{TiO}_{2}$ contents. The products were then severely deformed by twist extrusion technique. The microstructures, densities, wear resistances, hardness and strengths of the resulted composites, in two different $\mathrm{TiO}_{2}$ reinforcement content of 2 and $8 \mathrm{wt} \%$ and at various TE cycles were examined.
## 2. Experimental procedures
To attain a uniform distribution of the $\mathrm{TiO}_{2}$ reinforcement particles in $\mathrm{Al}-\mathrm{Cu}$ matrix, a high-energy planetary ball-mill machine, manufactured by the authors was utilized and, the powder behavior was studied during the process. Milling time and the effects of volume fraction of the reinforcement and its particle size were also investigated. Aluminum powder with mean grain size of $<45 \mu \mathrm{m}$ and commercial purity of $99.9 \%$ and copper powder of $40 \mu \mathrm{m}$ with $99.0 \%$ purity were supplied. Nano-scale anatase $\mathrm{TiO}_{2}$ powder, as the reinforcement, having a mean size of $50 \mathrm{~nm}$ was also obtained. The powders specifications are shown in Table 1.
The appropriate proportions of $\mathrm{Al}$ and $\mathrm{Cu}$ powders were weighed using a digital balance of $0.001 \mathrm{mg}$ accuracy. Internal surfaces of the cups were wetted with a thin layer of glycerin to prevent sticking the powders mixtures on them. The powders were then mixed and blended in a high energy planetary ball mill to produce the matrix alloy powder. For ball-milling, chromium steel balls with diameters of 17 , $19,22,25$ and $30 \mathrm{~mm}$, ball-to-powder weight ratio of $20: 1$, constant rotational speed of $300 \mathrm{rpm}$ and argon atmosphere were used. Ball milling time was $16 \mathrm{~h}$. To reinforce the product, $2 \mathrm{wt} \%$ and $8 \mathrm{wt} \%$ of $\mathrm{TiO}_{2}$ particles were added and blended to a homogeneous mixture. The mixture was cold compacted into a two-piece die of DIN-1.2344 hot die steel having a hole of $15 \times 15 \times 80 \mathrm{~mm}$ dimensions for $15 \mathrm{~min}$ under $600 \mathrm{MPa}$. Then, to enhance apparent densities of powders, they were put into the die under $100 \mathrm{MPa}$ pressure being heated to temperatures of $550^{\circ} \mathrm{C}, 580^{\circ} \mathrm{C}$ and $640^{\circ} \mathrm{C}$. After reaching these temperatures, the pressure was turned up to $700 \mathrm{MPa}$ at which the samples were kept for 30,60 and $120 \mathrm{~min}$ to obtain three different densities. After sintering, the samples were furnace cooled and homogenized to room temperature at a rate of $21.8^{\circ} \mathrm{C} / \mathrm{h}$.
For twist extrusion testing, the samples were lubricated with $\mathrm{MoS}_{2}$ to reduce friction. Then, they were inserted into the entrance guide of the twist extrusion die being pushed to the distorted channel using a steel plunger with speed of $1.1 \mathrm{~mm} / \mathrm{s}$. The twist extrusion die of $14.6 \times 14.6 \mathrm{~mm}$ internal cross-section with a twist line slope of $\beta=60^{\circ}$ in the counter-clockwise direction was used (Fig. 1). In order to apply a backpressure on the sample, the output channel was built steeped. This channel, itself, acted as a direct extrusion die. Thus, after the specimen passes the twisted channel, enters a straight output passage of $29 \mathrm{~mm}$ length during which its cross-section changed from $14.6 \times 14.6 \mathrm{~mm}$ to $14.2 \times 14.2 \mathrm{~mm}$. In addition, for preventing deviation of the sample to the sides and making sure of upright entering of the sample into the twisted channel, an $80 \mathrm{~mm}$ channel with $15 \times 15 \mathrm{~mm}$ crosssection was developed at the entrance as the sample guide. Also, to inhibit stress concentration, the right angle corners of the die interior walls were blunted. In this research, two sets of $\mathrm{Al}-4 \mathrm{wt} \% \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $\mathrm{Al}-4 \mathrm{wt} \% \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ samples were extruded at velocity of $68.4 \mathrm{~mm} / \mathrm{min}$ (maximum velocity of the available press) for 1,2 and 4 passes. The maximum allowed number of TE passes was found to be four, above which material failure occurred during twist extrusion operation.
To investigate the microstructure changes in the materials due to twist extrusion, the samples were prepared by cutting from the cross-section perpendicular to the axial direction of the extruded billets. The microstructure evolution was then studied in the central, lateral and corner regions of the cross-section using scanning electron microscopy (SEM).
Densities of the compacted powders were determined through Archimedes procedure according to the standard ASTM B93-13 [22]. The microstructures of samples from both composites were studied using SEM model ISI ABT SR-50 equipped with EDX analyzer after their preparation including grinding, polishing and etching with Keller etchant solution. To investigate the formation of deleterious phases like $\mathrm{Al}_{7} \mathrm{Cu}_{2} \mathrm{Fe}$ and $\mathrm{Al}_{4} \mathrm{C}_{3}, \mathrm{X}$-ray diffraction examinations and scanning electron microscopy observations were carried out on the sintered composites. To evaluate grain size and lattice strain, spectroscope system equipped with copper ray lamp (wavelength $1.5405 \AA$ ) was utilized. Williamson-Hall equation was used to determine crystallite size and lattice strain in diffracting domain. For hardness measurements of the sintered samples, Vickers hardness testing machine with the applied force of $1000 \mathrm{~g}$ was utilized.
To evaluate wear resistance of the composite specimens, tribological studies were conducted according to ASTM G99-04 standard [23] using a WAZAU pin-ondisk wear testing machine connected to computer interface from Tribo V4.3L software. The samples were cut from the cross-section perpendicular to the extrusion direction. Hardness testing was performed on the points across cross-section diameter with $1 \mathrm{~mm}$ intervals from one corner to the other. The samples were cut by a Merck lathe from the upper part of the cylindrical samples to make disks of $50 \mathrm{~mm}$ diameter and $4 \mathrm{~mm}$ thickness. To polish the sample surfaces, they were ground against 100,200 , and 500 grit emery papers. As the wearing apparatus, pins of $5 \mathrm{~mm}$ length and $2 \mathrm{~mm}$ diameter from 2160 steel with 60 HRC were prepared. The applied force and sliding distance were selected to be $30 \mathrm{~N}$ and $1000 \mathrm{~m}$, respectively. Wear coefficient, K, was predicted using the Archard equation [24]:
$V=K W L / H$
where $V$ is the lost volume of the worn material, $H$ Brinell hardness, $W$ normal applied load equal to $30 \mathrm{~N}, L$ the sliding distance (m) and $K$ wear coefficient.
Table 1
Specifications of powders used in this study.
| Powder | Particle size | Purity (\%) |
| :--- | :--- | :--- |
| $\mathrm{Al}$ | $<45 \mu \mathrm{m}$ | 99.9 |
| $\mathrm{Cu}$ | $<40 \mu \mathrm{m}$ | 99 |
| $\mathrm{TiO}_{2}$ | $<50 \mathrm{~nm}$ | 99 |
Fig. 1. Twist channel of TE die with $\alpha=90^{\circ}$ and $\beta=60^{\circ}$.
To investigate mechanical properties of the composites and to plot true stresstrue strain relation, compression tests were carried out. The cylindrical compression samples were cut from the centre of the billets for 1,2 and 4 cycles with ratio of $H / D=1$, separately out of the samples containing 2 and $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ and, then were prepared and polished. The tests were conducted at ambient temperature. True stress-true strain relation of samples in each pass were inferred from compressive stress-strain curves.
## 3. Results and discussion
### 3.1. Powders specifications
Firstly, to determine the adequate milling duration, aluminum powder and $4 \mathrm{wt} \% \mathrm{Cu}$ powder were mixed and ball-milled for 5, 6 and $8 \mathrm{~h}$. The XRD patterns are compared in Fig. 2a. As seen, with increasing the milling time, the XRD peak intensities for aluminum and copper phases decreased and the XRD peak intensity of $\mathrm{Al}_{2} \mathrm{Cu}$ phase increased, therefore, it is inferred that after $8 \mathrm{~h}$ milling, the phase $\mathrm{Al}_{2} \mathrm{Cu}$ was formed and alloying process was completed.Now, $2 \mathrm{wt} \% \mathrm{TiO}_{2}$ reinforcement powder was added to the mixture and milled for 4 more hours (two-stage alloying). Next, aluminum, copper and $\mathrm{TiO}_{2}$ powders were mixed together and were milled in two portions for 12 and $16 \mathrm{~h}$ (one-stage alloying). Compressive results are shown in Fig. 2b. In two-stage conditions, as observed in Table 2, the subgrain size is smaller, but formation of $\mathrm{Al}_{7} \mathrm{Cu}_{2} \mathrm{Fe}$ brittle phase occurred that might be due to gradual intrusion of $\mathrm{Fe}$ into the mixture during ball-milling through surface erosion of the balls and cups. On the other hand, the weak signs of the formation of that phase were observed in one-stage milling for $16 \mathrm{~h}$. Therefore, to minimize the possible formation of the deleterious brittle phase of $\mathrm{Al}_{7} \mathrm{Cu}_{2} \mathrm{Fe}$ in the final product, all samples were produced through one-stage, $16 \mathrm{~h}$ ball-milling. In Fig. 2c, the results of X-ray diffraction experiments for milling of $\mathrm{Al}-4 \mathrm{wt} \% \mathrm{Cu}$ powder mixture containing $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ for $16 \mathrm{~h}$ are depicted. The results indicated the formation of $\mathrm{Al}_{2} \mathrm{Cu}$ phase which witnesses that the alloying was performed successfully. The deleterious brittle phase of $\mathrm{Al}_{7} \mathrm{Cu}_{2} \mathrm{Fe}$ was not observed in the final product.
Fig. 2. XRD results (a) for different milling times of Al-4Cu, (b) comparison of results for $12 \mathrm{~h} 2$-stage with $12 \mathrm{~h}$ and $16 \mathrm{~h} 1$-stage mechanical alloying and (c) for $16 \mathrm{~h}$ milling of powder mixture containing $8 \mathrm{wt} \% \mathrm{TiO}_{2}$.
Table 2
A comparison of subgrain size and grain strain in different milling times.
| Sample | Type of milling | Milling time $(\mathrm{h})$ | Subgrain size $(\mathrm{nm})$ | Grain strain |
| :--- | :--- | :--- | :--- | :--- |
| $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ | Two stage | 12 | 11 | 0.0017 |
| $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ | One stage | 12 | 33 | 0.0045 |
| $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ | One stage | 16 | 31.51 | 0.00475 |
| $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{TiO}_{2}$ | One stage | 16 | 28.41 | 0.00481 |
Table 3
Relative densities of samples at 700 Mpa pressure for $30 \mathrm{~min}$ at different temperatures.
| Sample | Theoretical density $\left(\mathrm{g} / \mathrm{cm}^{3}\right)$ | Bulk density $\left(\mathrm{g} / \mathrm{cm}^{3}\right)$ | Temperature $\left({ }^{\circ} \mathrm{C}\right)$ | Relative density $(\%)$ | Porosity percentage $(\%)$ |
| :--- | :--- | :---: | :--- | :--- | :--- |
| $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ | 2.797 | $2.256 \pm 0.0006$ | 550 | 90.31 | 9.69 |
| | | $2.7095 \pm 0.0004$ | 580 | 96.87 | 3.13 |
| $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{TiO}_{2}$ | 2.856 | $2.7452 \pm 0.003$ | 640 | 98.15 | 1.85 |
| | | $2.759 \pm 0.004$ | 640 | 98.64 | 1.36 |
### 3.2. Evaluation of composites
For hot compacting, the twist extruded samples having $2 \mathrm{wt} \% \mathrm{TiO}_{2}$, first, temperature of $550^{\circ} \mathrm{C}$ and, then $580{ }^{\circ} \mathrm{C}$ were applied. Densities of those samples were calculated using dipping-in-water procedure (Eqs. (2) and (3)) and, since their measured densities at above two temperatures were found to be very low ( $<97 \%)$, higher temperatures were applied. The applied pressure for all samples was fixed at $700 \mathrm{MPa}$ for $30 \mathrm{~min}$. Porosity volume fraction was also determined using Eq. (3); the results are presented in Table 3. For applying a uniform axial pressure, the ratio of height to diameter $(h / d)$ was about 1.5 .
$\rho=\frac{W_{\text {air }}\left\lfloor\rho_{\text {water }}-0.0012\right\rfloor}{0.99983\left\lfloor W_{\text {air }}-W_{\text {water }}\right\rfloor}+0.0012$
$\rho_{T}=\sum_{i=1}^{n} f_{i} \rho_{i}$
$\%$ Porosity $=\frac{\rho_{T}-\rho}{\rho_{T}} \times 100$
where $W_{\text {air }}$ is the measured weight of the sample in air, $W_{\text {water }}$ the weight in water, $\rho$ measured density, $\rho_{\text {water }}$ density in water, $\rho_{T}$ theoretical density.In hot compression test performed on composite, temperature of $640^{\circ} \mathrm{C}$ was applied for the other samples, but their holding time in the furnace increased to 60 and $120 \mathrm{~min}$. The final densities are tabulated in Table 4. In hot compression situations at $640{ }^{\circ} \mathrm{C}$ for $120 \mathrm{~min}$, the density of the sintered sample was nearly equal to the theoretical density of the composite. Therefore, the same conditions were repeated for the mixture powder containing $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ reinforcement. Because the density of $\mathrm{TiO}_{2}$ particles was higher than that of the matrix alloy $\left(4.5 \mathrm{~g} / \mathrm{cm}^{3}\right)$, it was anticipated that with increasing volume fraction of reinforcement particles, the relative density of the composite increased [5] which was consistent with the results depicted in Table 4.
Fig. 4. SEM micrograph of sample $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ showing large agglomerated $\mathrm{TiO}_{2}$ particles.
Table 4
Relative densities of samples at $700 \mathrm{MPa}$ pressure at $640^{\circ} \mathrm{C}$ and different times.
| Sample | Theoretical density $\left(\mathrm{g} / \mathrm{cm}^{3}\right)$ | Bulk density $\left(\mathrm{g} / \mathrm{cm}^{3}\right)$ | Time $(\mathrm{min})$ | Relative density $(\%)$ | Porosity percentage $(\%)$ |
| :--- | :--- | :--- | :--- | :--- | :--- |
| $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ | 2.797 | $2.252 \pm 0.0008$ | 30 | 98.15 | 1.85 |
| | | $2.757 \pm 0.0005$ | 60 | 98.59 | 1.41 |
| $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{TiO}_{2}$ | 2.856 | $2.769 \pm 0.0003$ | 120 | 99.01 | 0.99 |
| | | $2.833 \pm 0.0004$ | 120 | 99.20 | 0.80 |
Fig. 3. SEM images of samples reinforced with (a) 2 and (b) $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ particles prior to $\mathrm{TE}$.
Fig. 5. (a and b) XRD patterns of composites reinforced with $2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ nano-particles and (c and d) EDX analysis results for those composites, respectively.
Table 5
Weight percentage of elements in $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ derived from EDX analysis.
| Element | wt\% of element in $\mathrm{Al}-4 \mathrm{Cu} /$ <br> $2 \mathrm{wt} \% \mathrm{TiO}_{2}$ | wt\% of element in $\mathrm{Al}-4 \mathrm{Cu} /$ <br> $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ |
| :---: | :---: | :---: |
| $\mathrm{Al}$ | 84.1 | 77.14 |
| $\mathrm{Cu}$ | 3.88 | 3.68 |
| $\mathrm{Ti}$ | 2.08 | 7.86 |
| 0 | 9.94 | 11.32 |
| Total | 100 | 100 |
Fig. 3 shows SEM micrographs of the samples containing $2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ prior to TE with coarse distinct grains. As seen, the reinforcement particles are distributed uniformly within the matrix. The particles became finer with smooth edges and corners during ball-milling. Uniform distribution of nanoparticles within matrix, because of their high surface to volume ratio, is difficult. In the composites with $8 \mathrm{wt} \% \mathrm{TiO}_{2}$, distribution of these nano-particles was inhomogeneous resulting in formation of their large agglomerates; these usually impair mechanical properties of materials. This is more evident in Fig. 4.
Fig. 6. (a) $\mathrm{SE}$ images of composite reinforced with $2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $\mathrm{EDX}$ analysis of point $\mathrm{A}$ and $\mathrm{B}$, (b) $\mathrm{SE}$ micrograph of composite reinforced with 8 wt $\% \mathrm{TiO}_{2}$ with $\mathrm{EDX}$ analysis of points $A$ and $B$.
Fig. 7. Compression true stress-true strain curves for annealed samples until beginning of barreling.
According to XRD results shown in Fig. 5a, presence of the brittle phase (i.e. $\mathrm{Al}_{7} \mathrm{Cu}_{2} \mathrm{Fe}$ in the sintered $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ sample, which causes brittle fracture of the material, is confirmed. However, X-ray diffraction pattern in Fig. $5 \mathrm{~b}$ indicates that in the $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{TiO}_{2}$ sample, the brittle phase is absent. In the next stages of experiments, it will be noticed that with application of severe plastic strains, fracture of latter samples occurred more frequently in comparison with the former ones. For performing quantitative analysis of the existing elements in the samples, EDX was utilized. The results are depicted in Fig. 5(c) and (d) as well as in Table 5. Those results were obtained from surfaces of the samples showing the total weight percentage of the elements in the sintered samples. The analysis showed no contamination.
Fig. 6(a) shows secondary electron (SE) micrograph of a sample with composition of $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ together with $\mathrm{EDX}$ analyses of points A and B (specified with circles), which are nearly identical. Therefore, it is concluded that the reinforcing particles are uniformly distributed within the matrix, as mentioned earlier. However, EDX analyses of points A and B of SE micrograph of the sample $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{TiO}_{2}$ shown in Fig. 6 (b) are quite dissimilar. It means that a non-uniform distribution of $\mathrm{TiO}_{2}$ nano-particles coupled with their large agglomerates, which usually contribute to deterioration of mechanical properties of the composite, occurred.
Compression true stress-true strain curves of the annealed samples, in the case of $\mathrm{Al}-4 \mathrm{Cu}, \mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $\mathrm{Al}-4 \mathrm{Cu} /$ $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ are illustrated and compared in Fig. 7 and the results are shown in Table 6. It is evident that with increasing $\mathrm{TiO}_{2}$ content in the composite, yield stress and Young modulus increase but, ductility decreases. Perhaps, deletion of porosities during hot compression at $640^{\circ} \mathrm{C}$ for $120 \mathrm{~min}$ was effective in enhancement of the sample strength. On the other hand, $\mathrm{TiO}_{2}$ particles are stable thermodynamically, and do not react with the matrix phase at high temperatures. These particles act as barriers against movement of dislocations leading to ductility decrease [25].
Fig. 8 shows variation of lost volume of composites during wear test vs. sliding distance. As previously shown, with increasing the percentage of $\mathrm{TiO}_{2}$ reinforcement, the lost volume decreases. To predict wear coefficient, Eq. (1) was used. The graph in Fig. 9 shows variation of wear coefficient with sliding distance for all samples. As indicated, the lowest wear coefficient belongs to the composite having $8 \mathrm{wt} \% \mathrm{TiO}_{2}$. It is also evident in Table 7 that with increasing weight percentage of $\mathrm{TiO}_{2}$ particles, the hardness of the material increases. Such a hardness increasing can be attributed to the increasing of dislocation densities improving material resistance. Hardness increasing leads to enhancement of wear resistance of the composite [25].
Fig. 8. Variation of lost volume of matrix alloy and composite vs. sliding distance in wear test.
Fig. 9. Variation of wear coefficient for matrix alloy and composites vs. sliding distance at pressure of $30 \mathrm{~N}$.
Table 7
Mean Brinell hardness for different samples.
| Specimen | Hardness (Brinell) |
| :--- | :--- |
| $\mathrm{Al}-4 \mathrm{Cu}$ | 107 |
| $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ | 138 |
| $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{TiO}_{2}$ | 200 |
Fig. 10 are the backscattered electron BSE images of samples $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ from central and lateral regions. EDX examinations revealed that the two discrete dark and light regions corresponding to $\mathrm{TiO}_{2}$ clusters and the matrix, respectively. Fig. 10(a), (b), (e) and (f) shows the microstructure of two composites in central area after two passes and four passes of twist extrusion, respectively. It is revealed that grains became finer and the microstructure was more uniform with increasing the number of TE passes and this is same for Fig. 10(c), (d), (g) and (h) that illustrate the microstructure of lateral region of cross-section of both composites after two passes and four passes of TE, respectively. After four passes of TE in both composites, formation of nano-sized grains are evident. Unlike significant effect of billet axial rotations between ECAP passes [26,27], the billet rotations between TE passes have no effect on the plastic flow. This is due to the axial symmetry of the process.
Table 6
Results of uniaxial compression testing for annealed samples.
| Specimen | Young modulus (GPa) | Yield strength (GPa) | Barreling stress (MPa) | Barreling strain (\%) |
| :--- | :--- | :--- | :--- | :--- |
| $\mathrm{Al}-4 \mathrm{Cu}$ | 66 | 256 | 270 | 0.96 |
| $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ | 70 | 246 | 0.93 | 284|
| $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{TiO}_{2}$ | 79.6 | 278 | 284 | 0.87 |
Fig. 10. BSC images of composites, up: $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and down: $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$, (a, b, e and f) central regions, after 2 and 4 passes, (c, d, $\mathrm{g}$ and $\mathrm{h}$ ) lateral regions, after 2 and 4 passes.
Fig. 11. Grain size measurement for a corner of cross-section of a sample containing $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ extruded for 2 passes.
Clustering of fine particles has been reported by Ritasalo et al. [28]. With increasing twist extrusion passes, $\mathrm{TiO}_{2}$ clusters became smaller having a more homogeneous distribution in the matrix. For the sample $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$, as observed, with increasing the number of TE passes, very fine equiaxed grains are formed. It is also evident that in the centre of the sample, the microstructure is less homogeneous and the formation of $\mathrm{TiO}_{2}$ clusters is observed. As indicated, in both types of samples, the microstructures in the corners and edges of the cross-section are more homogenous than those in the centers. This means that the corners experienced larger strains compared to the centre. In each pass of TE process, applying plastic deformation leads to an increase in dislocation density and, consequently formation of subgrains that impede dislocations glide gradually. With accumulation of plastic strains in subsequent passes, misorientation between neighboring grains increases and elongated grains turn to fine equiaxed ones resulting in a recrystallized microstructure [29]. This is the same phenomenon normally observed in SPD processes and is termed dynamic recrystallization [30].
Mishra et al. [29] suggested that when grains become smaller and also when the total area of grain boundaries increases, discrete dislocations emitted by a boundary are absorbed by the opposite grain boundary. Therefore, in higher passes of TE, dislocation density decreases gradually and high angle grain boundaries form. Shape of grains and rate of converting low angle boundaries to high angle ones in TE process depend on twist path and twist angle $(\alpha)$. It is worth noting that strain distribution and the boundary of deformation zones depend strongly on the geometry of die crosssection, i.e. deviation angle $(\beta)$ and twist angle $(\alpha)$, and by varying these factors, one can change strain intensity in different regions.
Grain sizes of the extruded samples were determined after each TE pass using a scaling-measuring utility installed to scanning electron microscope, as shown in Fig. 11 and the results are depicted in Table 8. It is inferred from Table 8 that with increasing TE passes, grain size in the centre and in the corners decreases. However, with increase of passes, the amount of strain that can be imposed on the sample, decreases. Such a decrease is further observed at edge regions. Therefore, uniformity of deformed structure increases and gradually, extent of grain refinement increases in central and lateral regions of the sample. This is due to structure stability brought about by saturation of the mechanical properties after the strain exceeds saturation limit. Such stability and saturation are not confined to TE, but are extended to all deformations based on pure shear like ECAP and so on. Mechanisms of this effect in PSD processes are such that with increasing passes (i.e. with increasing the strains), grain boundary surfaces also increase with a rate proportional to deformation state. During plastic deformation, cells or subgrains form and after a rather large strain, a considerable change does not occur in them. Therefore, with increasing strain, amount of high angle boundaries increases [29]. It is then concluded that with increasing the number of passes, the difference in grain sizes decreases in various regions of the sample. To determine extent of this difference in various passes, a variable index $(V)$ is defined as a ratio of standard deviation, SD to a parameter average value $\bar{x}$ as follows [31]:
$V=\frac{S D}{\bar{x}} \times 100$
Table 8
Mean grain size $(\mu \mathrm{m})$ of centers and corners of cross sections of two composites at various passes of TE.
| Sample | Position | Annealed | 1-pass | 2-pass | 4-pass |
| :--- | :--- | :--- | :--- | :--- | :--- |
| $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ | Corner | $20.9 \mu \mathrm{m}$ | $11 \mu \mathrm{m}$ | $9 \mu \mathrm{m}$ | $7.1 \mu \mathrm{m}$ |
| | Center | $21.7 \mu \mathrm{m}$ | $20 \mu \mathrm{m}$ | $12 \mu \mathrm{m}$ | $8.4 \mu \mathrm{m}$ |
| $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ | Corner | $18.2 \mu \mathrm{m}$ | $10 \mu \mathrm{m}$ | $7.8 \mu \mathrm{m}$ | $5.6 \mu \mathrm{m}$ |
| | Center | $18.8 \mu \mathrm{m}$ | $14 \mu \mathrm{m}$ | $9.8 \mu \mathrm{m}$ | $6.5 \mu \mathrm{m}$ |
Fig. 12. Grain size heterogeneity index for different passes for two composites.
Fig. 12 shows inhomogeneity of grain size for different passes and various $\mathrm{TiO}_{2}$ contents in composites. As indicated, the annealed sample is more homogeneous in grain size and, the sample extruded for one pass shows highest inhomogeneity. It is evident that with increasing the number of passes, the inhomogeneity of grain size decreases to a minimum of $V=\sim 16 \%$ for fourpass TE operation.
Fig. 13(a) shows the effect of $\mathrm{TiO}_{2}$ content on Vickers microhardness of the composites. A more uniform dispersion of $\mathrm{TiO}_{2}$ particle in the matrix impedes dislocation movements resulting in an increase of the hardness [32]. Fig. 13(b) and (c) shows the variation of hardness in the centre and corners of cross-section as well as mean hardness, in different passes of TE, for both types of composites. As seen, during first pass, the hardness increases dramatically, but in next passes, an obvious decrease in the curve slope is observed. This has been attributed to increase in dislocation density resulted from application of severe plastic deformation [29].Since corners get higher plastic strains than the centre, they possess higher hardness. However, with increasing the number of passes and gradual saturation of microstructure with strain due to saturation in dislocation density and, then development of a fine substructure, the heterogeneity in hardness distribution on the cross-section of the sample decreases. Therefore, despite occurrence of heterogeneity in deformation, hardness distribution is homogeneous at higher strains. Zendehdel et al. also reported homogeneity of hardness distribution at higher passes when they investigated influence of $\mathrm{TE}$ process on microstructure and mechanical properties of 6063 aluminum alloy [33]. Fig. 14(a) and (b) illustrates Vickers microhardness measured along diagonal line on cross-section of different samples. As indicated, hardness of samples increased noticeably after first pass compared to the annealed specimens. For the samples containing $2 \mathrm{wt} \% \mathrm{TiO}_{2}$, the hardness increased by $52 \%$ in average, but for the sample having $8 \mathrm{wt} \% \mathrm{TiO}_{2}$, the hardness increasing was $46 \%$. The hardness increase is lower for the central regions that undergoes lower strain levels and, higher for the lateral areas deformed by higher strains; this is due to higher redundant strains $\left(\varepsilon_{\mathrm{r}}\right)$ at the lateral regions [34]. The variation index, $V$, is also defined for description of heterogeneity level in hardness values. Using the variation index, $V$, calculated through Eq. (5) for different passes, Table 9 for the sample $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and Table 10 for the sample $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ are tabulated. Fig. 15 shows the heterogeneity index of hardness values $(V)$ for different passes of TE. It is inferred from Tables 9 and 10, and Fig. 15 that, in addition to increase in hardness within the central and lateral areas, and also increasing average hardness in whole sample, heterogeneity increased as well. Hardness heterogeneity index value in the sample of $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{TiO}_{2}$ extruded for 4 passes reached to 8.44 from 1.37 for the conditions before TE and, also in the sample of $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{TiO}_{2}$ extruded for 4 passes, $V$ value reached to 12.45 from 1.42 for the conditions before TE. It seems that heterogeneous distribution of hardness within whole sample was not significant due to application of backpressure during the process. Because backpressure is necessary for completion of cinematic conditions of applied plastic flow through tool geometry of $\mathrm{TE}$, and facilitates development of more homogeneous structure and mechanical properties [30].
Fig. 13. (a) Variation of Vickers hardness of composites with different $\mathrm{TiO}_{2}$ contents, prior to TE. (b) Variation of hardness at central and lateral regions of cross-section and mean hardness at different $\mathrm{TE}$ passes for composite $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$. (c) Variation of hardness at central and lateral regions of cross-section and mean hardness at different TE passes for composite $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$.
Fig. 14. Vickers hardness measured along diagonal line over cross section of (a) sample $\mathrm{Al}-4 \mathrm{Cu} 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ at various passes of $\mathrm{TE}$.
Fig. 15. Heterogeneity index of hardness values for various TE passes.
Table 9
Vickers microhardness values for samples $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ at different passes and their heterogeneity values.
| Sample | Hardness in center (HV) | Mean hardness (HV) | Hardness of edge (HV) | Heterogeneity V (\%) |
| :--- | :--- | :--- | :--- | :--- |
| Annealed | 145 | 146 | 147 | 1.3 |
| 1-pass | 217 | 224 | 231 | 6.25 |
| 2-pass | 225 | 234 | 243 | 7.69 |
| 4-pass | 227 | 237 | 247 | 8.44 |
Table 10
Vickers microhardness values for samples $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ at different passes and their heterogeneity values.
| Sample | Hardness in center (HV) | Mean hardness (HV) | Hardness of edge (HV) | Heterogeneity V (\%) |
| :--- | :--- | :--- | :--- | ---: |
| Annealed | 210 | 211.5 | 213 | 1.42 |
| 1-pass | 296 | 308.5 | 321 | 8.10 |
| 2-pass | 318 | 335.5 | 353 | 10.43 |
| 4-pass | 324 | 345.5 | 367 | 12.45 |
Fig. 16. Results of compression tests: (a) true stress-true strain curves for $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$, (b) true stress-true strain curves for $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO} \mathrm{O}_{2}$. (c) Variation of yield strength at different passes for both composites.
Table 11
Results of uniaxial compression testing on samples $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $\mathrm{Al}-4 \mathrm{Cu} / 8 \mathrm{wt} \% \mathrm{TiO}_{2}$ (Type 1 and Type 2, respectively) at different passes.
<table><thead><tr><th rowspan="2">Sample</th><th colspan="2">Young modulus (GPa)</th><th colspan="2"> Yield strength (MPa)</th><th colspan="2">Barreling strain MPa</th><th colspan="2"> Barreling strain (%)</th></tr><tr><th>Type 1</th><th>Type 2</th><th>Type 1</th><th>Type 2</th><th>Type 1</th><th>Type 2</th><th>Type 1</th><th>Type 2</th></tr></thead><tr><td>Annealed</td><td>70</td><td>79.6</td><td>246</td><td>278</td><td>284</td><td>298</td><td> 0.93</td><td> 0.87</td></tr><tr><td>1-pass</td><td>84</td><td> 94.9</td><td> 304</td><td> 325</td><td>241</td><td>347</td><td>0.80</td><td>0.74</td></tr><tr><td>2-pass</td><td>85</td><td>96.6</td><td>334</td><td>354</td><td>358</td><td>366</td><td>0.66</td><td>0.62</td></tr><tr><td>4-pass</td><td>87.6</td><td>97.2</td><td>368</td><td>372</td><td>383</td><td>391</td><td>0.65</td><td>0.61</td></tr></table>
Fig. 16(a) and (b) shows graphs obtained from compression tests including true stres-true strain curves derived from uniaxial compression test on samples $\mathrm{Al}-4 \mathrm{Cu} / 2 \mathrm{wt} \% \mathrm{TiO}_{2}$ and $\mathrm{Al}-4 \mathrm{Cu} /$ $8 \mathrm{wt} \% \mathrm{TiO}_{2}$ after different TE passes, respectively. The obtained results are compared in Table 11. Fig. 16(c) illustrates variation curves of yield strength for both composites in different TE passes. As indicated, with increasing number of passes, strength increased and ductility decreased. The strength increase extent in the first pass is obviously higher than that in the second and fourth passes which is due to gradual strain saturation in different regions, particularly those closer to the centre. By applying a few number of TE passes, strain exceeds saturation limit and saturation state gradually extends all cross-section area leading to uniformity in changing microstructure and other properties [35].
## 4. Conclusions
Mechanical alloying and powder metallurgy routes were applied to fabricate a new composite with binary matrix of Al$4 \mathrm{Cu}$ and reinforced by $2 \mathrm{wt} \% \mathrm{TiO}_{2}$ nano-particles. The annealed materials were subjected to some of the mechanical tests, and hardness, strength and yield strength were measured. With increasing percentages of the reinforcement particles, hardness, yield strength, Young modulus and wear resistance of the composites increased but ductility decreased. Afterwards, the composites were deformed severely through twist extrusion for 1,2 and 4 passes. The maximum allowable passes of extrusion was four, beyond which the materials did not endure plastic deformation and failed. It was revealed that with increasing the number of passes by 4 , a more homogeneous distribution of reinforcement particles occurred and also an ultrafine-grained nano-structure was obtained.
## References
[1] B. Ogel, R. Gurbuz, Microstructural characterization and tensile properties of hot pressed Al-SiC composites prepared from pure Al and Cu powders, Mater. Sci. Eng. A 301 (2001) 213-220.
[2] J. Onoro, M.D. Salvador, L.E.G. Cambronero, High-temperature mechanica properties of aluminium alloys reinforced with boron carbide particles, Mater. Sci. Eng. A 499 (2009) 421-426.
[3] R. Khorshidi, A. Hassani, Comparative analysis between TOPSIS and PSI methods of materials selection to achieve a desirable combination of strength and workability in Al/SiC composite, Mater. Des. 52 (2013) 999-1010.
[4] A. Hassani, M. Zabihi, High strain rate superplasticity in a nano-structured Al$\mathrm{Mg} / \mathrm{SiCP}$ composite severely deformed by equal channel angular extrusion, J. Mater. Des. 39 (2012) 140-150
[5] H. Kaftelena, M.L. Ovecoglua, H. Heneinb, H. Cimenoglua, $\mathrm{ZrC}$ particle reinforced $\mathrm{Al}-4 \mathrm{wt} \% \mathrm{Cu}$ alloy composites fabricated by mechanical alloying and vacuum hot pressing: microstructural evaluation and mechanical properties, Mater. Sci. Eng. A 527 (2010) 5930-5938,
[6] S.A.A. Akbari Mousavi, A.R. Shahab, M. Mastoori, Computational study of Ti$6 \mathrm{Al}-4 \mathrm{~V}$ flow behaviors during the twist extrusion process, Mater. Des. 29 (2008) 1316-1329.
[7] R.Z. Valiev, Y. Estrin, Z. Horita, T.G. Langdon, M.J. Zehetbauer, Y.T. Zhu, Producing bulk ultrafine-grained materials by severe plastic deformation, JOM 58 (4) (2006) 33-39.
[8] N.A. Smirnova, V.I. Levit, V.I. Pilyugin, R.I. Kuznetsov, L.S. Davydova, V.A. Sazonova, Evolution of structure of fcc single crystals during strong plastic deformation, Phys. Met. Metallogr. 61 (6) (1989) 127-134.
[9] M.V. Segal, V.I. Reznikov, A.E. Drobyshevskiy, V.I. Kopylov, Plastic metal working by simple shear, Izvestia Akademii nauk SSSR. Metally 1 (1981) 115123.
[10] G.A. Salishchev, O.R. Valiakhmetov, R.M. Galeyev, Formation of submicrocrystalline structure in the titanium alloy VT8 and its influence on mechanical properties, J. Mater. Sci. 28 (1993) 2898-2903.
[11] Y. Saito, H. Utsunomiya, N. Tsuji, T. Sakai, Novel ultra-high straining process for bulk materials development of the accumulative roll-bonding (ARB) process, Acta Mater. 47 (2) (1999) 579-583.
[12] J.Y. Huang, Y.T. Zhu, H.G. Jiang, T.C. Lowe, Microstructures and dislocation configurations in nanostructured Cu processed by repetitive corrugation and straightening, Acta Mater. 49 (9) (2001) 1497-1505.
[13] Y. Beygelzimer, V.N. Varyukhin, D.V. Orlov, S.G. Son, Twist extrusion - the accumulation of strain, Donetsk firma naukoemnih Technol. Natl. Acad. Sci. Ukraine (2003) 73-75.
[14] N. Pardis, R. Ebrahimi, Deformation behavior in simple shear extrusion (SSE) as a new severe plastic deformation technique, Mater. Sci. Eng. A 527 (1-2) (2009) 355-360.
[15] V. Varyukhin, Y. Beygelzimer, R. Kulagin, O. Prokofeva, A. Reshetov, Twist extrusion: fundamentals and applications, Mater. Sci. Forum 667-669 (2011) 31-37.
[16] Y. Beygelzimer, D. Orlov, V. Varyukhin, Proceedings of the Second International Symposium on Ultrafine Grained Materials, in: Y.T. Zhu (Ed.), Minerals, Metals, and Materials Society, Warren Dale (PA), 2002, pp. 297-301.
[17] Beygelzimer Y, Varukhin V, Synkov S, Sapronov A, Synkov V. New techniques for accumulating large plastic deformations using hydroextrusion, Fizika i Tekhnika Vusokih Davlenii (High Pressure Physics and Technology, in Russian) 1999; 9(3).
[18] Y. Beygelzimer, V. Varyukhin, D. Orlov, S. Sinkov, Twist extrusion: accumulating deformations 56 (2003) 456-465 (in Russian).
[19] Y. Beygelzimer, D. Prilepoa, R. Kulagina, V. Grishaeva, A. Abramovaa, V. Varyukhina, Planar twist extrusion vs. TWIST extrusion, J. Mater. Process Technol. 211 (2011) 522-529.
[20] V. Varukhin, Y. Beygelzimer, S. Synkov, D. Orlov, Applications of twist extrusion, Mater. Sci. Forum 503-504 (2006) 335-340.
[21] Y. Beygelzimer, V. Varyukhin, S. Synkov, D. Orlov, Useful properties of twist extrusion, Mater. Sci. Eng. A 503 (2009) 14-17.
[22] ASTM B962-13 Standard Test Methods for Density of Compacted or Sintered Powder Metallurgy (PM) Products Using Archimedes' Principle West Conshohocken, PA: ASTM International, 2004.
[23] G99-04 A. Standard test method for wear testing with a Pin-on-Disk apparatus. West Conshohocken, PA: ASTM International, 2004.
[24] J.F. Archard, Contact and rubbing of flat surfaces, J. Appl. Phys. 24 (1953) 981.
[25] C.S. Ramesha, A.R. Anwar Khanb, N. Ravikumar, P. Savanprabhu, Prediction of wear coefficient of $\mathrm{Al}_{6061}-\mathrm{TiO}_{2}$ composites Wear 259 (2005) 602-608.
[26] M. Furukawa, Y. Iwahashi, Z. Horita, M. Nemoto, T.G. Langdon, The shearing characteristics associated with equal-channel angular pressing, Mater. Sci. Eng. A 257 (2) (1998) 328-332.
[27] M.V. Segal, Slip line solutions, deformation mode and loading history during equal channel angular extrusion, Mater. Sci. Eng. A 271 (1-2) (1999) 322-333.
[28] R. Ritasalo, M.E. Cura, X.W. Liu, Y. Ge, T. Kosonen, U. Kanerva, O. Söderberg, S.P. Hannula, Microstructural and mechanical characteristics of $\mathrm{Cu}_{-1} \mathrm{Cu}_{2} \mathrm{O}$ composites compacted with pulsed electric current sintering and hot isostatic pressing, Composites: Part A 45 (2013) 61-69.
[29] A. Mishra, V. Richard, F. Gregori, R.J. Asaro, M.A. Meyers, Microstructural evolution in copper processed by severe plastic deformation, Mater. Sci. Eng. A 290 (2005) 410-411.
[30] D. Orlov, Y. Beygelzimer, S. Synkov, V. Varyukhin, N. Tsuji, Z. Horita, Plastic flow, structure and mechanical properties in pure Al deformed by twist extrusion, Mater. Sci. Eng. A 519 (2009) 105-111.
[31] Y. Beygelzimer, D. Orlov, A. Korshunov, S. Synkov, V. Varyukhin, I. Vedernikova, A. Reshetov, A. Synkov, L. Polyakov, I. Korotchenkova, Features of twist extrusion: method, structures and material properties, Solid State Phenomena 114 (2006) 69-78.
[32] G.S. Kataiah, D.P. Girish, The mechanical properties and fractography of aluminium $6061-\mathrm{TiO}_{2}$ composites, IJPSR I (2010) 17-25.
[33] H. Zendehdel, A. Hassani, Influence of twist extrusion process on microstructure and mechanical properties of 6063 aluminum alloy, Mater. Des. 37 (2012) 13-18.
[34] M.S. Mohebbi, A. Akbarzadeh, Experimental study and FEM analysis of redundant strains in flow forming of tubes, J. Mater. Process Technol. 210 (2010) 389-395
[35] Y. Beygelzimer, A. Reshetov, S. Synkov, O. Prokofeva, R. Kulagin, Kinematics of metal flow during twist extrusion investigated with a new experimental method, J. Mater. Proc. Technol. 209 (2009) 3650-3656.
\ No newline at end of file
tests/benchmark/datasets/annotations/cleaned/cleaned_academic_literature_f7904bc37cc2e25c1e3e412978854b10.md 0 → 100644
View file @ 09117e54
# Research Article
## The effects of AZD3582 [4-(nitroxy)butyl-(2S)-2-(6-methoxy-2naphthyl) propanoate], and naproxen on key pathogenic steps in NSAID-enteropathy in the rat.
M. Walley, G. Sigthorsson, C. Hotz-Behofsits, R. Simpson, I. Bjarnason*
Guy's, King's and St Thomas' School of Medicine, Department of Gastroenterology, King's College Hospital Foundation Trust, Denmark Hill,<br>London SE5 9PJ, London, UK, Tel: ++2032992417, Fax: ++2032996474, e-mail: ingvar.Bjarnason @kcl.ac.uk
Received 9 October 2006; revised 27 January 2007; accepted 1 February 2007
Abstract. Background: The pathogenesis of NSAID-induced enteropathy may involve dual inhibition of the cyclooxygenase (1 and 2 ) and a topical effect with sequential increased intestinal permeability, development of inflammation and ulcers. It has been suggested that nitric-oxide donating drugs cause significantly less gastrointestinal injury by counteracting for NSAID-induced reductions in blood flow.
Aims: To compare the effects of AZD3582 [4-(nitroxy)butyl(2S)-2-(6-methoxy-2-naphthyl) propanoate], and naproxen on key pathogenic steps in NSAID-enteropathy in the rat.
Methods: Single doses of AZD3582, naproxen (dose range $10-300 \mu \mathrm{mol} / \mathrm{kg}$ ) or vehicle were given to male Sprague Dawley rats. Intestinal permeability ( ${ }^{1} \mathrm{CrEDTA}$ ) and intestinal inflammation (granulocyte marker protein) was quantitated and ulcer counts made.
Results: Intestinal permeability (all doses) and inflammation (highest dose of the drugs) increased significantly from control levels following naproxen and AZD3582 and there was no significant difference between the drugs. Median ulcer counts were, however, significantly ( $\mathrm{p}<0.01$ ) lower with AZD3582 (4 $\pm 2$ ) than with naproxen $(17 \pm 4)$.
Conclusions: Naproxen and AZD3582 are equally associated with increased small intestinal permeability and inflammation, which is the consequence of their topical effect. The reduced small bowel ulcer counts with AZD3582 accords with the suggestion that vascular factors are the main driving force for NSAID-induced ulcer formation.
Key words: NSAIDs; CINODs; Nitric oxide; Naproxen; AZD3582[^0]
## Introduction
Nonsteroidal anti-inflammatory drugs (NSAIDs) cause gastrointestinal side effects which involve the stomach as well as the small bowel mucosa (Hawkey and Langman, 2003). Although the serious gastric side effects of bleeding and perforation have attracted the most attention, it is increasingly clear that the small bowel is associated with similar types and prevalence of complications (Bjarnason et al., 1993; Laine et al., 2002)
The pathogenesis of NSAID-induced gastrointestinal damage is uncertain. There is substantial evidence to suggest that NSAID-enteropathy in rodents is caused by various combinations of the selective biochemical effects of NSAIDs including cyclooxygenase (COX)-1 and COX-2 inhibition together with the topical effect (Somasundaram et al., 1995). The topical effect is thought to relate to the physicochemical properties of NSAIDs to act as detergents (Lichtenberger et al., 1995) and uncouplers of mitochondrial oxidative phosphorylation (Somasundaram et al., 1995). Collectively, it is suggested that the topical effect results in increased intestinal permeability with mucosal exposure of luminal aggressive factors and hence inflammation. It is also suggested that NSAID-induced inhibition of COX1, with decreased amounts of vasoactive prostaglandins, drives this inflammation to ulcers (Wallace et al., 2000). Interestingly small bowel damage occurs with long-term COX-2 absence or inhibition (small bowel inflammation and ulcers) although the mechanisms are unclear (Sigthorsson et al., 2002).
One of the more recent suggestions for reducing the intestinal side effects of NSAIDs is to attach a nitric oxide (NO) moiety to the NSAID, in the hope that the NO might counteract the effect of prostaglandin deficiency on the intestinal microcirculation. NO donors have been used in patients with cardiovascular disease for more than a century (Burgaud et al., 2002). NO is recognised as an important modulator of a large number of physiological processes (Wallace and Miller, 2000). More specifically, NO increases mucosal blood flow and mucus secretion and decreases leukocyte adherence (Cirino et al., 1996; Wallace and Miller, 2000). As these actions could, in theory, counteract the effects of mucosal prostaglandin deficiency induced by NSAIDs, a class of COX-inhibiting nitric oxide donors (CINODs) have been developed. CINODs are frequently produced by the addition of a nitroxybutyl moiety to the carboxylic group of the NSAID (which mediates the binding to the COX enzymes) by means of an ester linkage (Fiorucci, 2001). As well as potentially offering improved gastrointestinal tolerability, CINODs may result in enhanced anti-inflammatory, anti-pyretic and analgesic effects when compared to NSAIDs (Fiorucci et al., 2002) or COX-2 selective agents, although this is a controversial issue.
Previous animal studies have indicated that CINODs may offer reduced adverse gastrointestinal effects when compared to the parent compounds (Elliott et al., 1995; Davies et al., 1997). More recently, human studies have demonstrated that NO-aspirin (NCX-4016) maintains COX-1 and platelet inhibition whilst nearly avoiding the short-term gastric damage (Fiorucci, et al., 2003) and that AZD3582 [4-(nitroxy)butyl(2S)-2-(6-methoxy-2-naphthyl) propanoate] reduces gastrointestinal toxicity when compared to naproxen (Hawkey et al., 2003).
The precise reason that these CINODs reduce the gastrointestinal damage of the parent drug is controversial. Their ester linkage to the NSAID abolishes their topical effect and their effect on the COX enzymes. In order to maintain therapeutic efficacy the CINOD needs to undergo hydrolysis yielding the parent NSAID and the NO moiety. If the beneficial effects of CINODs are due to their counteracting or compensating for the vascular effects of prostaglandin deficiency it might be expected that CINODs would be equally associated with the permeability and inflammatory changes of the comparator NSAID (consequence of the topical effect) whilst reducing the number of ulcers. We tested this hypothesis by comparing the effects of AZD3582 with those of naproxen on small bowel permeability (using ${ }^{51} \mathrm{CrEDTA}$ ), intestinal inflammation (quantitated by measurement of granulocyte marker protein (GMP)) and small bowel ulcer counts in rats.
## Methods
## Animals
Male Sprague Dawley rats (Charles Rivers), 6-8 weeks old, weighing $200-250 \mathrm{~g}$ were used throughout these studies. Two groups of animals were used. The first group was used for the measurement of intestinal permeability and ulcer counts while the second group was used to assess intestinal inflammation via the measurement of granulocyte marker protein (GMP) in stool samples. Animals were housed singly in metabolism cages for up to 9 days and fed standard laboratory diet and water. For measurement of intestinal inflammation, stool samples were collected each day from day 2 to day 9 (with the drugs being administered on day 5). Following an overnight fast (day 4), animals received AZD3582 or naproxen diluted in solvent (oil in water emulsion) from $60 \mu \mathrm{mol} / \mathrm{ml}$ emulsions and given by gastric gavage at the doses of $10,30,100$ or $300 \mu \mathrm{mol} / \mathrm{kg}$ ( $\mathrm{n}=8$ in each group). Control animals received solvent only. Ulcer counts were performed 48 hours after administration of the drugs or vehicle. The naproxen and AZD3582 were obtained from AstraZeneca, R\&D Sodertalje, Sweden.
## Intestinal Permeability
One hour after administration of the study drugs or vehicle, the rats were administered $10 \mu \mathrm{Ci}{ }^{51} \mathrm{CrEDTA}$ via a tube in a volume of $0.5 \mathrm{ml}$ water followed by $1 \mathrm{ml}$ of water. Animals were allowed food and fluids 2 hours later. All urine passed during the following 5 hours was collected and the samples were assayed for gamma-radioactivity along with standards ( $10 \%$ of the dose given) in a Wallac 1284 gamma counter (Pharmacia, Sweden) for 1 minute. Results are expressed as percentage of the oral dose that was excreted in urine, which provides a measure of intestinal permeability as previously described (Somasundaram et al., 2000).
## Intestinal Inflammation
Stool samples were collected on each day of the study and $1 \mathrm{~g}$ wet weight of each were added to $4 \mathrm{ml}$ of extraction buffer (Tris $50 \mathrm{mM}$, $\mathrm{NaCl} 150 \mathrm{mM}, \mathrm{CaCl}_{2} 10 \mathrm{mM}$, Thiomersal $0.25 \mathrm{mM}, \mathrm{pH}$ to 8.4$)$. The samples were then homogenized for 30 seconds at $20,000 \mathrm{rpm}$ using an Ultra Turrax homogenizer (IKE Werke, Germany) and spun in a microcentrifuge for 10 minutes at $13,000 \mathrm{rpm}$. The supernatant was decanted off into an eppendorf tube and the samples were assayed for GMP as previously described (Sigthorsson et al., 2002). In short $50 \mu 1$ of a 1:200 dilution in duplicate to 96 well microtitre plates were added. The plates were pre-coated with anti-GMP antibody. Equal volumes of 9 standards were also added to the plates in duplicate. The plates were incubated at room temperature on a plate shaker for 45 minutes, washed 4 times with rinsing buffer (Tris $50 \mathrm{mM}, \mathrm{NaCl} 150 \mathrm{mM}, \mathrm{MgCl}_{2} 0.5 \mathrm{mM}, \mathrm{KCl}$ $2.5 \mathrm{mM}$, Thiomersal $0.25 \mathrm{mM}$, Tween- $200.05 \%, \mathrm{pH}$ to 8.0 ) allowed to dry and then $50 \mu \mathrm{l}$ of alkaline phosphatase (ALP) conjugated anti-GMP (diluted 1:800 in assay buffer) was added to each well. The plates were incubated under the same conditions as before, washed and dried as before and then $100 \mu \mathrm{l}$ of substrate (p-nitrophenyl phosphate, $1 \mathrm{mg} / \mathrm{ml}$, in substrate buffer ( $10 \%$ diethanolamine), $\mathrm{MgCl}_{2} 0.5 \mathrm{mM}$, Thiomersal $0.25 \mathrm{mM}, \mathrm{pH} 9.6$ ) was added to each well. The optical density of the highest standard was monitored and when it read between 1.2-1.8, the reaction was stopped by adding $50 \mu 11 \mathrm{M} \mathrm{NaOH}$ to each well. The plates were read at $405 \mathrm{~nm}$ using an MRX plate reader plus Dynex Revelation software (Dynex Technologies, USA). The results are expressed in $\mathrm{mg} / \mathrm{l}$ of extract.
## Macroscopic studies
To assess ulceration within the small bowel, animals were euthanazed by $\mathrm{CO}_{2}$ inhalation 48 hours after administration of the drugs. The abdomen was opened via a midline incision and the small intestine isolated, removed and gently flushed with $0.9 \%$ saline. The intestinal mucosa was exposed by cutting along the anti-mesenteric side of the intestine. Ulcer counts were performed by noting both the number and size of the ulcers ( $\leq 5 \mathrm{~mm}$ were recorded as pointed, $>5 \mathrm{~mm}$ were recorded as longitudinal).
## Statistical analysis
Results are presented as median and range as not all data was normally distributed. Wilcoxon's rank sum test was used to assess statistical differences between groups and the Wilcoxon's signed rank test for sequential data.
## Results
## Intestinal Permeability
Figure 1 shows that administration of both naproxen and AZD3582 significantly increased intestinal permeability at all doses given when compared with baseline (vehicle only) ( $p<0.001$ ). There was no significant difference ( $p>0.05$ ) in intestinal permeability between AZD3582 and naproxen at any of the doses given.
## Intestinal Inflammation
There was no significant increase in GMP with the 10 or $30 \mu \mathrm{mol} / \mathrm{kg}$ doses of either drug (Figure 2). Rats given AZD3582 at a dose of $100 \mu \mathrm{mol} / \mathrm{kg}$ had GMP values significantly higher than the control group ( $p<0.05$ ). At doses of $300 \mu \mathrm{mol} / \mathrm{kg}$, a significant increase in intestinal inflammation was noted with both naproxen and AZD3582 when compared to the vehicle group ( $\mathrm{p}<0.01$ ). No significant difference was observed between the two drugs at any of the dose range tested.
## Macroscopic examination
On macroscopic examination, no ulcers were seen with either drug over a dose range of $0-100 \mu \mathrm{mol} / \mathrm{kg}$. The mean number of ulcers with naproxen $300 \mu \mathrm{mol} / \mathrm{kg}$ ) was 17.1 (range 10 29). The rats treated with AZD3582 had significantly fewer ulcers ( $\mathrm{p}<0.001$ ) (median 2.5; range 0-12) (Figure 3).
Fig. 1 Urinary excretion of 51 CrEDTA after Naproxen and AZD3582. The white circles represent median (bars represent range) values obtained from rats dosed with naproxen. The black circles represent values obtained from rats dosed with AZD3582. Urinary excretion of $51 \mathrm{CrEDTA}$ was measured 5 hours following dosing.
Fig. 2 GMP concentrations after Naproxen and AZD3582. The white circles represent median (bars represent range) GMP values obtained from rats dosed with naproxen. The black circles represent the GMP values obtained from rats dosed with AZD3582. Data shown is taken from the day following dosing.
Fig. 3 Small bowel ulcer counts after Naproxen and AZD3582. The white circles represent the number of ulcers in rats dosed with naproxen. The black circles represent the number of ulcers in rats dosed with AZD3582. Counts were made 48 hours after dosing with $300 \mu \mathrm{mol} / \mathrm{kg}$.
## Discussion
These studies show that AZD3582 is associated with significantly less small bowel ulcerative damage than naproxen while the postulated consequences of the topical effect, intestinal permeability and inflammation, is equally evident with both drugs. The findings are consistent (assuming that the NO is released prior to or during drug absorption ren-dering intact naproxen) with the aforementioned pathogenic framework for NSAID-induced small bowel damage and the suggestion that the NO maintains vascular perfusion following the administration of naproxen.
A number of studies show that virtually all acidic NSAIDs increase small intestinal permeability, by virtue of their acidity and lipid solubility, and it is suggested that that this is a prerequisite for the development of small intestinal inflammation (Sigthorsson et al., 2000). Unlike non-selective NSAIDs, the NO moiety of AZD3582 renders the molecule non-acidic, and hence it can not exert a topical effect in this form. Nevertheless it is still associated with increased intestinal permeability in the current experiments suggesting hydrolyses of the ester bond, presumably by gastric and more importantly pancreatic esterases (Somasundaram et al., 1997). This raises the possibility that the beneficial action of AZD3582 on the stomach (Hawkey et al., 2003; WilderSmith et al., 2006) may simply be due to its lack of topical toxicity (Rainsford and Whitehouse, 1980).
While both drugs were associated with similar increases in intestinal permeability there were no inflammatory changes following the lower ( 10 or $30 \mu \mathrm{mol} / \mathrm{kg}$ ) doses of the drugs unlike previous studies where inflammation invariably follows the intestinal permeability changes (Somasundaram et al., 1997; Somasundaram et al., 2000). The reasons for this may be that much higher doses of NSAIDs were administered in previous studies. Indeed at the higher doses of 100 and $300 \mu \mathrm{mol} / \mathrm{kg}$, dose dependent inflammation was seen for naproxen and AZD3582, the inflammation being similar for both drugs. It is noteworthy that NO itself, despite its potentially beneficial effect on microvascular blood flow and healing, may be directly toxic to the epithelial cells at high concentrations (Menconi et al., 1998).
The results of studies on AZD3582 can not be extrapolated over to other NO-NSAIDs as their method of production, stability, pharmacokinetics and rate of hydrolyses may differ. A similar study to the current one (Davies et al., 1997) nevertheless found more contrasting degrees of inflammation with naproxen compared to the CINOD. This study used higher drug doses and a much stricter dosing regime (twice daily dosing for over 2 weeks compared to our single dose over 8 days). However, Somasundaram (1997) using nitroxybutyl-flurbiprtofen obtained almost identical results to the current study.
Previous studies have clearly dissociated the consequences of the topical effect (increased intestinal permeability and inflammation) from the COX-1 inhibitory effect, which seems to drive the inflamed mucosa to an ulcerated one (Somasundaram et al., 2000). The ulcerative damage with AZD3582 was significantly less than for naproxen. The precise mechanism for this is nevertheless uncertain. Estimates of the metabolism of CINODs and NO-releasing drugs suggest that the rate of NO release from these compounds both in vitro and in vivo is slow in comparison to other NO donors such as sodium nitroprusside (SNP) and S-nitroso-N-acetyl-D,L-penicillamine (Keeble and Moore, 2002). However NO certainly has the potential to increase microvascular blood flow (Whittle, 2003) and thus reduce the damage (Wallace et al., 2000), but it also increases mucous secretion, reduces secretion and adhesion of neu- trophils and reduces cytokine release from macrophages, all of which may be impaired by COX inhibition (Wallace et al., 2000). It has also been suggested that cytochrome P450 may play a role in the metabolism of CINODs (Grosser and Schroder, 2000). Alternatively the bioavailability of naproxen from a dose of AZD3582 may be lower than from a dose of naproxen.
In summary, AZD3582 is associated with equal changes in increased intestinal permeability and inflammation as equimolar doses of naproxen. At the same time it is associated with significantly less ulcerative small bowel damage. These findings are consistent to the suggestions that NO derived from AZD3582 counteracts the vascular effects of NSAID-induced inhibition of COX.
Acknowledgements. The drugs and solvents were supplied by AstraZeneca, Sweden who supported this project.
## References
Bjarnason, I., Hayllar, J., Macpherson, A. J. et al. (1993). Side effects of nonsteroidal anti-inflammatory drugs on the small and large intestine in humans. Gastroenterology 104: 1832-1847.
Burgaud, J. L., Ongini, E. and Del Soldato, P. (2002). Nitric oxidereleasing drugs - A novel class of effective and safe therapeutic agents. Ann N. Y. Acad. Sci. 962: 360-371.
Cirino, G., Wheeler-Jones, C. P., Wallace, J. L. et al. (1996). Inhibition of inducible nitric oxide synthase expression by novel nonsteroidal anti-inflammatory derivatives with gastrointestinal-sparing properties. Br. J. of Pharmacol. 117: 1421-1426.
Davies, N. M., Roseth, A. G., Appleyard, C. B. et al. (1997). NO-naproxen vs naproxen: Ulcerogenic, analgesic and anti- inflammatory effects. Aliment. Pharmacol. Ther. 11: 69-79.
Elliott, S. N., McKnight, W., Cirino, G. et al. (1995). A Nitric Oxide-Releasing Nonsteroidal Antiinflammatory Drug Accelerates GastricUlcer Healing in Rats. Gastroenterology 109: 524-530.
Fiorucci, S. (2001). NO-releasing NSAIDs are caspase inhibitors. Trends in Immunology 22: 232-235.
Fiorucci, S., Antonelli, E., Mencarelli, A. et al. (2002). A NO-releasing derivative of acetaminophen spares the liver by acting at several checkpoints in the Fas pathway. Br. J. of Pharmacol. 135: 589-599.
Fiorucci, S., Santucci, L., Greasele, P. et al. (2003). Gastrointestinal safety of NO-aspirin (NCX-4016) in healthy human volunteers: a proof of concept endoscopic study. Gastroenterology 124: 600-607.
Grosser, N. and Schroder, H. (2000). A common pathway for nitric oxide release from NO-aspirin and glyceryl trinitrate. Biochem. Biophys. Res. Commun. 274: 255-258.
Hawkey, C. J., Jones, I. J., Atherton, C. T. et al. (2003). Gastrointestinal safety of AZD3582, a cyclooxygenase inhibiting nitric oxide donator: proof of concept study in humans. Gut 52: 1537-1542.
Hawkey, C. J. and Langman, M. J. (2003). Non-steroidal anti-inflammatory drugs: overall risks and management. Complementary roles for COX-2 inhibitors and proton pump inhibitors. Gut 52: 600-608.
Keeble, J. E. and Moore, P. K. (2002). Pharmacology and potential therapeutic applications of nitric oxide-releasing non-steroidal anti-inflammatory and related nitric oxide-donating drugs. $B r . J$. of Pharmacol. 137: 295-310.
Laine, L., Bombardier, C., Hawkey, C. J. et al. (2002). Stratifying the risk of NSAID-related upper gastrointestinal clinical events: Results of a double-blind outcomes study in patients with rheumatoid arthritis. Gastroenterology 123: 1006-1012.
Lichtenberger, L. M., Wang, Z-M., Romero. J. J. et al. (1995) Non-steroidal anti-inflammatory drugs (NSAIDs) associate with zwitterionic phospholipids: Insight into the mechanism and reversal of NSAID-induced gastrointestinal injury. Nat Med 1: 154-158
Menconi, M. J., Unno, N., Smith, M. et al. (1998). Nitric oxide donor-induced hyperpermeability of cultured intestinal epithelial monolayers: role of superoxide radical, hydroxyl radical, and peroxynitrite. Biochim. Biophys. Acta. 1425: 189-203.
Rainsford, K. D. and Whitehouse, M. W. (1980). Anti-inflammatory antipyretic salicylic acid esters, with low gastric ulcerogenic activity. Agents Actions. 10: 451-456.
Sigthorsson, G., Tibble, J., Mahmud, T. et al. (2000). NSAID-Induced gastrointestinal damage: the biochemical consequences of the "ion trapping" hypothesis. Inflammopharmacology 8: 31-41.
Sigthorsson, G., Simpson, R. J., Walley, M. et al. (2002). COX-1 and 2 , intestinal integrity and pathogenesis of NSAID-enteropathy in mice. Gastroenterology 122: 1913-1923.
Somasundaram, S., Hayllar, J., Rafi S. et al. (1995). The biochemical basis of NSAID-induced damage to the gastrointestinal tract: A review and a hypothesis. Scand J Gastroenterology. 30: 289-299.
Somasundaram, S., Rafi, S., Jacob, M. et al. (1997). Intestinal tolerability of nitroxybutyl-flurbiprofen in rats. Gut 40: 608-613.
Somasundaram, S., Sigthorsson, G., Simpson, R. J. et al. (2000). Uncoupling of intestinal mitochondrial oxidative phosphorylation and inhibition of cyclooxygenase are required for the development of NSAID-enteropathy in the rat. Aliment Pharmacol Ther 14: 639650.
Wallace, J. L. and Miller, M. J. S. (2000). Nitric oxide in mucosal defense: A little goes a long way. Gastroenterology 119: 512-520.
Wallace, J. L., McKnight, W., Reuter B. K. et al. (2000). NSAID-induced gastric damage in rats: requirement for inhibition of both cyclooxygenase 1 and 2. Gastroenterology. 119: 706-714.
Wilder-Smith CH, J. B., Fornstedt-Wallin, B., Hedman, A. et al. (2006). Dose-effect comparisons of the CINOD AZD3582 and naproxen on upper gastrointestinal tract mucosal injury in healthy subjects. Scand J Gastroenterol 41: 264-273.
Whittle, B. J. (2003). Nitric oxide and the gut injury induced by nonsteroidalanti-inflammatory drugs. Inflammopharmacology 11:415422.
tests/benchmark/datasets/annotations/cleaned/cleaned_academic_literature_fbdb99151e811688574c0c4c67341074.md 0 → 100644
View file @ 09117e54
# Artificial Intelligence for 6G Networks: Technology Advancement and Standardization
Muhammad K. Shehzad, Luca Rose, M. Majid Butt, István Z. Kovács, Mohamad Assaad, and Mohsen Guizani
Abstract—With the deployment of 5G networks, standards organizations have started working on the design phase for sixth-generation ( $6 \mathrm{G}$ ) networks. $6 \mathrm{G}$ networks will be immensely complex, requiring more deployment time, cost and management efforts. On the other hand, mobile network operators demand these networks to be intelligent, self-organizing, and cost-effective to reduce operating expenses (OPEX). Machine learning (ML), a branch of artificial intelligence (AI), is the answer to many of these challenges providing pragmatic solutions, which can entirely change the future of wireless network technologies. By using some case study examples, we briefly examine the most compelling problems, particularly at the physical (PHY) and link layers in cellular networks where ML can bring significant gains. We also review standardization activities in relation to the use of ML in wireless networks and future timeline on readiness of standardization bodies to adapt to these changes. Finally, we highlight major issues in ML use in the wireless technology, and provide potential directions to mitigate some of them in $6 \mathrm{G}$ wireless networks.
Index Terms-AI, ML, Wireless networks, 3GPP, 6G.
## I. INTRODUCTION
Unprecedented growth in the global cellular traffic (as shown in Fig. 1) and immense data rate demands have become a challenge, leading wireless industry to the next-generation, called 6G. 6G-era will bring digital, physical and biological worlds together with the goal to improve human experience and well-being. $6 \mathrm{G}$ will be operating in TeraHertz $(\mathrm{THz})$ frequencies $(0.1-10 \mathrm{THz})$, hence beneficial for multiple use cases in industrial applications, providing immense data rates $(\approx 1 \mathrm{~Tb} / \mathrm{s})$, accelerating internet-of-things, and wider network coverage. AI/ML will pave the way for $\mathrm{THz}$ communications at different layers [2], e.g., supporting channel acquisition [3] and modulation classification [4] at PHY. Similarly, at the link layer, beamforming design and channel allocation can exploit ML [2]. In $\mathrm{THz}$ systems, a channel can significantly vary at a micrometer scale, resulting in a tremendous increase in channel estimation frequency and corresponding overhead. ML algorithms can counter this issue by using, e.g., improved channel prediction techniques [3], [5].
Fig. 1. Estimation of global mobile subscriptions in machine-to-machine (M2M) and mobile broadband (MBB) from 2020 to 2030. Source: ITU-R Report M. $2370-0$ [1].
Recently, fast-growing deployment of $5 \mathrm{G}$ has opened up many challenges, including massive complexity in network architecture, low latency, high cost, power consumption, and deployment of hybrid Long-Term Evolution (LTE) new radio $(\mathrm{NR})$, leading to difficulties in network optimization. In such a complex scenario, the network intelligence has become a major focus as it will play a pivotal role in complex problem solving [6], e.g., self-healing, self-optimization, and self-configuration of a network [7].
Future networks will become "cognitive" in a way that many aspects such as spectrum sensing/sharing, slicing, radio resource management (RRM), and mobility management, will be ML-based. Further, it is expected that ML will impact 6G air interface fundamentally and it will be designed to support ML natively [8]. Several recent research attempts, e.g., [9], propose different road maps for 6G, but they do not address standardization timeline and related issues regarding application of ML in 6G. Albeit, to some extent, [10] gives an overview of ML and standardization; nevertheless, ML-related technical challenges and its applications from an industrial and standardization perspective are not addressed.
Reconfigurable intelligent surface (RIS) and non-orthogonal multiple access (NOMA) are two key technologies for 6G [11]. RIS can re-engineer electromagnetic waves, hence beneficial to deliver the information where obstacles block the destination. RIS can be integrated with ML, allowing RIS to acquire envi-ronmental information by configuring various sensors, while ML can learn dynamic parameters intelligently, reducing the computation cost of RIS-based networks. Similarly, NOMA is a promising access technique for $6 \mathrm{G}$. In ML-empowered NOMA-based networks, gNodeBs ( $\mathrm{gNB}$ ) can intelligently define their control policy and improve decision-making ability.
Fig. 2. An overview of ML paradigms, major tools, and applications in wireless networks.
Today's networks use model-based methods to optimize various network functions providing characteristics of the process involved. However, these models might be too complex to be implemented in a realistic time frame or they include a great level of abstraction to function in a general environment. In contrast, ML-based solutions can adapt to real-time (RT) scenario changes and localized characteristics, learning the specific environment around the transceivers. The contributions of this article are twofold:
- We look at the above-mentioned problems from an industrial perspective and outline the gap between research and practice.
- We review standardization activities in the context of adopting ML in various aspects of wireless communications, e.g., channel acquisition, positioning. Furthermore, we highlight major issues and possible research directions in relation to the use of ML in wireless networks.
## II. OVERVIEW OF ML TECHNIQUES IN WIRELESS NETWORKS
ML is a process of training machines through data without explicit programming. Broadly speaking, ML consists of three paradigms: unsupervised learning, supervised learning, and reinforcement learning (RL). All these paradigms have a training/exploration phase to optimize a learning algorithm that later can be used in prediction/exploitation phase to infer on unknown inputs. As shown in Fig. 2, we briefly summarize them by providing some use cases in wireless networks.
1) Supervised Learning: Supervised learning exploits a labelled data set to learn a (hidden) function that maps an input to an expected output based on the examples. The standard techniques used to solve supervised learning-based problems are artificial neural networks (ANNs), support vector machines (SVMs), Bayesian networks, recurrent neural networks (RNNs), and convolutional neural networks (CNNs).
2) Unsupervised Learning: Unsupervised learning does not learn from labelled data, instead, training is based on an unlabelled data set. K-means and principal component analysis (PCA) are examples of two major tools used for clustering and dimensionality reduction, respectively.
3) Reinforcement Learning: RL is not based on training but rather the agent/decision-maker learns and decides online, maximizing a long-term reward. RL is beneficial in control problems where the agent adapts to changing environmental conditions, e.g., uplink power control.
Motivated by the considerable benefits of ML in various fields, its applications have also been considered in wireless networks almost at all layers of communication. Here, we focus on its impact on radio access networks (RAN), particularly PHY and link layers. Based on ML tools, given in Fig.2, some case studies will be explained later in Section III.
## A. Machine Learning at PHY
At PHY, many optimization problems are non-convex, e.g., sum-rate maximization. ML is a powerful tool to find good solution(s) for such non-convex optimization problems. Based on advanced learning algorithms, 6G networks provide the following major advantages by using ML.
- ML can be effective to deal with network complexity. 6G networks will be more complex due to numerous network topologies, immense growth in the cellular users, staggering data rate demands, complex air interface, vast network coordination methods, etc. Forecasting considerable complexity of $6 \mathrm{G}$ networks, the derivation of optimum performance solutions is nearly infeasible without ML.
- ML can play a vital role to deal with model deficit problems. Current cellular networks are amenable for mathematical derivation, for instance, information theory gives closed-form expressions for various problems such as Shannon theorem. However, the inherent complexity of $6 \mathrm{G}$ networks hinders the possibility of exploiting closed-form analytical expression(s), which can be due, for instance, to non-linearities either in the channel or network devices. ML offers an efficient way to deal with non-linearities, providing feasible solution(s) in a tractable manner.
- ML can cope with algorithm deficit problems. In current cellular networks, many optimal algorithms, although well-characterized, are impractical to be implemented. Considering the example of multiple-input multipleoutput (MIMO) systems where optimal solutions are known (e.g., dirty paper coding), they are overlooked in favour of linear solutions, e.g., linear minimum meansquared error. It is envisaged that ML can pave the way to implement more efficient yet practical solutions.
ML has been used to study various PHY issues, and without being exhaustive, some of the recent areas include:
- CNNs are used for modulation classification in [4].
- An RNN-based wireless channel predictor [5] is used in [3], explained in Section III-C to deal with inaccurate channel state information (CSI).
## III. Wireless Networks: Case Studies
In this section, we present three use cases to demonstrate the use of ML techniques in industrial wireless networks. ML tools utilized for these use cases are depicted in Fig. 2.
## A. UE Positioning
Highly accurate user equipment (UE) positioning is one of the prime considerations for Third Generation Partnership Project (3GPP) studies beyond Release 15. Various angle and time-of-arrival-based methods are used to determine UE positioning in today's cellular networks. All of these methods require triangulation techniques to resolve UE position and suffer from time synchronization errors.
We studied UE position by using radio frequency (RF) fingerprinting and two ML techniques, namely deep learning and decision tree, for an outdoor scenario [12]. Serving cell Reference Signal Received Power (RSRP) as well as neighbor cell RSRP values were used as features to train a deep neural network (DNN). As shown in Fig. 3, nearly $5 \mathrm{~m}$ accuracy is achieved for DNN when only 4 serving cell RSRP values and corresponding beam IDs are considered as a feature input, while it improves to nearly $1 \mathrm{~m}$ when 2 more RSRP values from the strongest neighboring cells, respective cell and beam IDs are added to the input feature set. The decision tree, a less complex algorithm as compared to DNN, provides about $2 \mathrm{~m}$ accuracy using data from both serving and neighboring cell beams as an input feature. The mean accuracy of nearly $1 \mathrm{~m}$ obtained from DNN is comparable to the accuracy level achieved with traditional methods without requiring triangulation and does not suffer from signal timing synchronization issues.
## B. ML-Assisted Proactive Mobility
For seamless and efficient mobility, a well optimized network should reduce the number of Handover (HO) events while avoiding Handover Failures (HOF) and Radio Link Failures (RLF). An emerging approach is to utilize ML-based algorithms, which enable proactive and UE specific mobility actions in the gNB. A relatively simple approach to this is to design an ML-based estimator of the radio measurements, such as RSRP of serving and neighbor cells, with a certain minimum accuracy and within a certain time horizon. Radio measurements are traditionally performed at the UEs side and reported to the serving $\mathrm{gNB}$ (or gNB-Centralized Unit) according to specific Radio Resource Control (RRC) configurations. For ML-based prediction purposes, time-traces of RSRP, or Reference Signal Received Quality (RSRQ) values need to be collected either in the UE and/or serving the gNB.
Fig. 3. Comparison of UE position for both DNN and decision tree techniques. The system level parameters for the network includes 8 sites with Inter-site distance $110 \mathrm{~m}$ and carrier frequency $28 \mathrm{GHz}$. For details of the parameters, please refer to [12].
For example, collected time-series of RSRP values are used as input to the ML-based predictor, which provides at the UE, and/or at the serving $\mathrm{gNB}$, a set of sufficiently accurately estimated RSRP values within a given future time horizon. Then, these signal estimations are used for predictive evaluation of possible $\mathrm{HO}$ conditions, thus can trigger proactive measurement reports from the UE and/or proactive $\mathrm{HO}$ actions at the serving $\mathrm{gNB}$. These two steps are repeated with a time periodicity given, e.g., by the sampling rate and time filtering of the input RSRP measurements [13], or alternatively, the steps can also be triggered by the serving $\mathrm{gNB}$ when certain traffic or mobility Quality-of-Service (QoS) conditions are met.
The outlined ML-based mobility algorithm can be implemented in either the UE or gNB or both, depending on the available ML assistance capabilities in each node. Furthermore, the mechanism can be integrated in self-organizing network-based Mobility Robustness Optimization solutions.
## C. CSI Feedback
CSI feedback in the downlink channel is a major challenge in Release 17 and beyond. Currently, CSI precision is affected by compressing the measurements imposed by the standard.
In our study, summarized in Section II-A, we assumed two RNN-based twin channel predictors at the $\mathrm{gNB}$ and UE [3]. The past CSI is utilized for training the RNN at both ends of the communication system. UE's feedback is evaluated with respect to the predicted channel. Fig. 4 depicts the meansquared error (MSE) between the actual channel versus the acquired channel at the $\mathrm{gNB}$ and the precoding gain when different quantization bits are used to feedback the CSI from the UE. The results are compared with and without using ML for the CSI feedback. A clear benefit of using ML can be observed. We believe that ML-based solutions will improve current performance without increasing signaling overhead.
(a) Trend of MSE.
(b) Trend of precoding gain.
Fig. 4. Performance of MSE and precoding gain. $2 \times 1$ MIMO configuration is considered, and RNN is composed of 1 hidden layer. For parameters' details, refer to [3].
## IV. Role of ML in Standardization
The potential of ML for $5 \mathrm{G}$ has been widely acknowledged in the literature and applications made it even in the standard at higher levels, e.g., for networking and security [7]. 3GPP has introduced a specification, named network data analytics function (NWDAF), in Release 15 and 16, as part of the $5 \mathrm{G}$ Core $(5 \mathrm{GC})$ architecture [7]. NWDAF is responsible for providing network analytics when requested by a network function (NF). Data is collected via application function (AF), operation, administration, and maintenance (OAM), NF, and data repositories. The specifications have also addressed the problem of inter-working for automation and data collection, which analytics vendors previously faced. 3GPP NWDAF framework for $5 \mathrm{G}$ systems is depicted in Fig.55. This automation gives leverage to network vendors for the deployment and testing of non-RT ML-related use cases. In Fig. 5. inward interfaces aggregate data from different network sources, where communication occurs using existing service-based interfaces. Outward interfaces provide decisions (analytics-based, algorithmic) to AF and NF.
Fig. 5. A generalized framework for 5G network automation in Release 16, representing that NWDAF should be able to collect data from the operator OAM, AFs and $5 \mathrm{GC}$ network functions $[7]$.
Regarding PHY, ML techniques lag behind, due to a number of issues. First, PHY makes use of abstractions and mathematical models that are inferred from the physical reality and electromagnetic principles. As long as such models describe the real-world precisely, there is no need for ML. Nevertheless, in practice, models and fixed algorithms are inefficient when facing rapidly changing and heterogeneous environments. For example, using the same channel acquisition scheme to acquire CSI from a laptop in line-of-sight with a $\mathrm{gNB}$, a tablet on a fast train, or a mobile quickly moving in a super densely covered area might not be optimal. Consequently, the standardization efforts of intelligent techniques have gained momentum, and while 3GPP is ready to begin a study item on ML implementations, open-radio access network (O-RAN) will be ML-native, defining a RAN intelligent controller (RIC), which will enhance several RAN functions.
3GPP has started studying the implications of the ML use at layer-1 and a study item on ML for NR air interface has been agreed upon. After the RAN-1 working group studies, protocol aspects will be studied in RAN-2 and subsequently, interoperability and testability aspects will be considered in RAN-4 working group. The remaining part of this section summarizes the status of the standardization of ML techniques for PHY for both 3GPP and O-RAN.
## A. CSI Feedback
CSI feedback for downlink channel in Release 17 is a complex issue in which UE-based beam selection is followed by CSI reference symbols (RS) training and precoding matrix index (PMI) reporting, and lastly by Demodulation Reference Signal (DMRS) and consequent estimation of the precoded channel. Broadly, beam selection aims to establish a sufficiently strong link budget between the UEs and the gNB. The CSI-RS is used for fine channel estimation, which is then fed back to the gNB to compute a precoder (eventually multiuser); finally, DMRS are precoded pilots that the UEs use to implement coherent demodulation. Currently, each of these phases is created following pre-established rules, with little to none room for intelligent behaviour. ML has been envisioned to possibly enhance each phase in a different way. Beam selection can be improved by intelligently correlating the beams with position or identity of the UEs. This would allow for a smart selection of the beams from the gNB side, thus avoiding brute-force selection. The CSI-RS can be enhanced by compressing the pilots and the PMI feedback exploiting ad hoc ML compressors. Furthermore, channel prediction techniques [5] can be used in order to pre-establish a baseline for the CSI feedback [3]. Other aspects that can be improved include frequency of pilots in both CSI-RS and DMRS, power and timing and CSI-RS port selection.
## B. $R S-D M R S$
Roughly speaking, DMRS are RS used for channel estimation to perform coherent demodulation. The correct estimation of the channel using such pilots have a strong impact on the performance in terms of bit-error-rate and thus block-errorrate. The role of the ML in such domain is twofold. First, it can be used to improve the performance of the channel estimation. Second, the ML can provide a smarter positioning of DMRS in order to reduce their frequency, hence reducing the overhead footprint in $6 \mathrm{G}$.
## C. Positioning
A precise positioning is one of the aspects that sees the largest improvement with respect to LTE's observed time difference of arrival (OTDOA) and uplink time difference of arrival (UTDOA), defined in Release 9 onward. Various aspects of $6 \mathrm{G}$ allow for precise positioning of the UE, such as large number of antenna elements at the $\mathrm{gNB}$, millimeter wave transmissions, dense network deployment. However, the methods based on angle-of-arrival and time-of-arrival fall short when non-line-of-sight scenarios are considered, in interference-limited scenarios. ML techniques, see Fig.2, are expected to help in improving the position by exploiting channel charting, hence learning the likely position of a UE based on a report, and multiplexing together information that carries positioning information but are hard to exploit in a classical way, such as CSI report and sounding reference signal maps.
## D. Mobility Enhancements
In 6G, frequent cell-selection, and frequent RSRP measurement could impact UEs' battery life. Furthermore, load balancing algorithms can use intelligent techniques that exploit the UE specific channel prediction, movement trajectory prediction and traffic demands prediction. Furthermore, the scenarios like fast-trains or non-terrestrial networks, will pose challenges to $\mathrm{HO}$ and conditional-HO operations. Novel solutions envisaged, compared to current 3GPP Release 17, include the use of UE specific ML-based predictive algorithms, addressed in Section III-B, designed to reduce paging errors and HO failures; thus, improve the overall QoS.
## E. Standardization for ML Data Collection
3GPP has started working on data collection for running ML algorithms in 5G networks [14]. The scope of such studies include identifying mechanisms to collect data from the network through minimization of drive test framework or further advanced enhancements. Furthermore, studies will focus on discussing hosting of ML models both for training as well as inference purposes at various network entities for various use cases and defining any new interfaces required for transporting data to the models.
## F. Federated Learning Model Collection
Training and prediction based on ML models will put an extra load on networks already transporting a large volume of data. Therefore, it is important to estimate the effect of model training and inference on network traffic, particularly for federated learning (FL) where UEs will act as distributed hosts [15]. The latency in collecting locally trained models is bounded in FL and network links should be able to meet delay budgets. This is particularly challenging in today's networks where a UE's own QoS requirements are already demanding and the FL model training and collection will further incur an extra burden on the network. Similarly, the split inference, where UEs cooperate with each other to perform joint inference, results in increasing the network traffic. 3GPP studies in Release 18 [15] will focus on the above mentioned issues to support training and inference for ML/FL models over wireless links.
## G. O-RAN-RIC
O-RAN alliance, aims to define a RAN network that is non-vendor specific, and that has an innate support for ML as an enabler for automation and OPEX savings. O-RAN alliance has defined interfaces for exchange of information in the protocol stack. To this end, in the O-RAN architecture, ML-assisted RAN intelligent controller (RIC) is included for network automation, for both scenarios, i.e., non-RT and RT. In the non-RT RIC, ML algorithms' training is done by using the data obtained at lower layers. However, the learning process remains slow; therefore, it is called non-RT RIC. Later, the learner is fed into the RT RIC, which utilizes the RT captured data to perform decisions online. Additionally, the functionality of non-RT includes policy management and higher layer procedure optimization. Therefore, the RAN or core-network can deploy such a mechanism based on the collected data.
## V. Open Challenges and Roadmap for Deploying ML TECHNIQUES
Though ML is a potential technology and enabler for nextgeneration wireless networks, several challenges related to its practical use are addressed below.
## A. Data Availability and Benchmarking
One of the foremost challenges in wireless networks is data availability. Data availability concerns the problem of identifying a common and accepted set of data (e.g., channel realizations) with the goal of testing and benchmarking ML algorithms. Such a problem is of a pivotal importance for standardization, where normally algorithms and proposals are tested using agreed underlying physical models (e.g., urban macrocells/microcells channel models), evaluation methodologies and calibrated simulators. Contrary to other fields, cellular networks have no standard data set to train and benchmark an ML algorithm. Therefore, a synthetic data set or software generated data set is of a predominant importance to train and benchmark ML algorithm(s), and to agree on a common evaluation methodology to rank proposition and standard algorithms.
Identifying a set of key performance indicators in wireless networks is another crucial task for ML standardization. It is necessary to design a set of metrics to classify and rank ML algorithms and their performance. Classic approaches such as throughput and signal-to-interference-plus-noise ratio (SINR) might not be sufficient since a small improvement in these values might come at the cost of large complexity augmentation and exacerbated energy consumption.
Fig. 6. Model collection for FL in a wireless network when some of the UEs have large blockage and use D2D communication for model transfer. Cluster-based UE selection is another solution for asynchronous model collection to meet network QoS requirements.
## B. Selection of ML versus Non-ML Solutions
ML tools are regarded as an implementation-oriented tool rather than a standard relevant aspect. The idea behind this relies on the fact that each vendor has the freedom to efficiently implement each aspect of the standard as long as the external interfaces are respected. A simple example of this is given in the CSI feedback, where a UE needs to select a specific PMI, but the standard does not specify any specific way in which this selection is performed. Recently, however, the idea of having ML dedicated message exchanges and performance that only an ML-aided algorithm can achieve has paved the way for standardization of ML algorithms [3]. This opens the door for several issues, e.g., will the standard impose a specific ML structure, classifying minimum performance and implementation structure, or will it remain far from the implementation? With regards to NNs, it is still open if hyperparameters are going to be left to vendor-specific implementation or will they be set by the standard.
## C. Complexity of ML Algorithms
Considering the limited battery life, storage, computational capability, and limited communication bandwidth in most cellular network entities, an ML model's cost-performance tradeoff becomes a fundamental issue. Another issue is the speed/time-steps at which the training and inference needs to be performed. Whereas hard-wired gNB have sufficient computational power to run complex ML algorithms, UEs need to face battery, heating and stringent complexity limits. Possible solutions to such issue include, but not limited to implementation of substitute rule-based algorithms at the UE side, migrating the load all on the $\mathrm{gNB}$ side.
## D. Communication-aware Federated Learning
Traditional ML models support centralized learning. Due to difficulties in collecting large amount of training data from the UEs, privacy issues and bandwidth bottleneck, FL has emerged as a promising solution. In FL, training is performed distributively over network devices, called local model hosts, and an application server on the network side acts as a central host to aggregate local models transmitted by the local learners. Typically, an application server host aggregates models only when updates are available from all the local learners, called synchronous model transfer. However, this is highly inefficient in wireless networks where links are unpredictable, local learners (UEs) are energy limited and have their own QoS requirements. Asynchronous model collection is the most viable solution for FL in wireless networks, where a subset of UEs is selected for a local model update in each round of model collection. However, UE selection in each round is a complex problem because UEs are energy limited and the network bandwidth is scarce, hindering collection of local models from all the UEs to represent independently and identically data collection. These mechanisms are usually vendor proprietary, but standardization still needs to define some common mechanisms for efficient model collection. As shown in Fig. 6. UE clustering and local device-to-device (D2D) communication for asynchronous model collection are possible solutions to decrease network communication and will require standardization support.
## E. Stability and Adaptability of ML Techniques
ML algorithms applied to wireless networks must be adaptive as they will have to deal with parameters that change dynamically. Particularly, the weights of the NN are evaluated online based on the trained data. However, this approach may not be applicable in wireless, and specifically in a standard, where coordination among entities belonging to different operators and provided by different vendors have to coexist, and in which the need for quick response could prevent one or the other solution. Possible solutions include: pre-trained $\mathrm{NN}$, or partially trained $\mathrm{NN}$ (i.e., $\mathrm{NN}$ in which the starting point is pre-set); cloud-based downloadable data set for $\mathrm{NN}$ training; codebook-based $\mathrm{NN}$, in which a codebook of different NNs is used and agreed upon between the gNB and UEs. Another related problem is to detect an outdated ML model with high inference error and replace it. Replacing an outdated model with a new model incurs further delay. Thus, there must be a proactive mechanism to adapt the ML model to network conditions such that network functions suffer minimum performance loss.
## VI. Conclusion
Motivated by the promise of the use of ML algorithms, we presented an overview of ML techniques to be used in 5G-Advanced and 6G wireless networks. Furthermore, we discussed the key roles of ML-based solutions from industrial and standardization perspectives. We also highlighted the practical challenges of deploying ML techniques in wireless networks and how to deal with them. Non-RT and higher layer ML-based solutions can be, and are, applied already in today's networks. Implementing RT ML solutions at PHY/MAC in 6G networks are the next big challenge in the research community. We believe that overcoming these challenges, both in research as well as at standardization levels, will pave the way for next-generation wireless communication to be effective and sustainable.
## REFERENCES
[1] I. Union, "IMT traffic estimates for the years 2020 to 2030," Report ITU, pp. 2370-0, 2015.
[2] A.-A. A. Boulogeorgos, E. Yaqub, M. Di Renzo, A. Alexiou, R. Desai, and R. Klinkenberg, "Machine learning: A catalyst for $\mathrm{THz}$ wireless networks," Frontiers in Communications and Networks, p. 37, 2021.
[3] M. K. Shehzad, L. Rose, and M. Assaad, "Dealing with CSI compression to reduce losses and overhead: An artificial intelligence approach," in 2021 IEEE International Conference on Communications Workshops (ICC Workshops), 2021, pp. 1-6.
[4] T. O'Shea and J. Hoydis, "An introduction to deep learning for the physical layer," IEEE Transactions on Cognitive Communications and Networking, vol. 3, no. 4, pp. 563-575, 2017.
[5] M. K. Shehzad, L. Rose, S. Wesemann, and M. Assaad, "ML-based massive MIMO channel prediction: Does it work on real-world data?" IEEE Wireless Communications Letters, pp. 1-5, 2022.
[6] B. Mao, F. Tang, Y. Kawamoto, and N. Kato, "Optimizing computation offloading in satellite-UAV-served 6G IoT: A deep learning approach," IEEE Network, vol. 35, no. 4, pp. 102-108, 2021.
[7] 3GPP, "Study of enablers for network automation for 5G (Release 16)," https://portal.3gpp.org/desktopmodules/Specifications/ SpecificationDetails.aspx?specificationId=3252, , Technical Report (TR) $23.791,062019$.
[8] J. Hoydis, F. A. Aoudia, A. Valcarce, and H. Viswanathan, "Toward a 6G AI-native air interface," IEEE Communications Magazine, vol. 59, no. 5, pp. 76-81, 2021.
[9] F. Tariq, M. R. Khandaker, K.-K. Wong, M. A. Imran, M. Bennis, and M. Debbah, "A speculative study on 6G," IEEE Wireless Communications, vol. 27, no. 4, pp. 118-125, 2020.
[10] R. Shafin, L. Liu, V. Chandrasekhar, H. Chen, J. Reed, and J. C. Zhang, "Artificial intelligence-enabled cellular networks: A critical path to beyond-5G and 6G," IEEE Wireless Communications, vol. 27, no. 2, pp. 212-217, 2020.
[11] R. Zhong, Y. Liu, X. Mu, Y. Chen, and L. Song, "AI empowered RISassisted NOMA networks: Deep learning or reinforcement learning?" IEEE Journal on Selected Areas in Communications, vol. 40, no. 1, pp. $182-196,2022$.
[12] M. M. Butt, A. Pantelidou, and I. Z. Kovács, "ML-assisted UE positioning: performance analysis and 5G architecture enhancements," IEEE Open Journal of Vehicular Technology, vol. 2, pp. 377-388, 2021.
[13] 3GPP, "NR; Radio Resource Control (RRC); Protocol specification (Release 15)," https://portal.3gpp.org/desktopmodules/Specifications/ SpecificationDetails.aspx?specificationId=3197 , Technical report (TR) TS38.331, 032021.
[14] - , "Study on enhancement for data collection for NR and ENDC (Release 17)," https://portal.3gpp.org/desktopmodules/Specifications/ SpecificationDetails.aspx?specificationId=3817 , Technical report (TR) $37.817,012021$.
[15] -, "5G System (5GS); Study on traffic characteristics and performance requirements for AI/ML model transfer (Release 18)," https://portal.3gpp.org/desktopmodules/Specifications/ SpecificationDetails.aspx?specificationId=3721 , Technical report (TR) $22.874,032021$.
Muhammad K. Shehzad [S'21] is working as a Research Engineer and Ph.D. student at Nokia Bell-Labs and CentraleSupelec, Paris, France, respectively. He received his B.Eng. (Hons.) degree in Electrical and Electronic Engineering from the University of Bradford, Bradford, U.K., in 2016, and M.S. in Electrical Engineering from the National University of Sciences \& Technology (NUST), Islamabad, Pakistan, in 2019. His major research interest is in MIMO communication using Artificial Intelligence (AI)/Machine Learning (ML).
Luca Rose [M'11] is Senior research and standard-ization expert with Nokia Bell-labs. He received his M.Sc. from university of Pisa, Italy, and his Ph.D. in Physics from Centrale-Supelec. He worked with Huawei France research center and Thales Communications and Security, contributing to several standard organizations. He is currently an ITU-R and ETSI delegate and the lead editor of IEEE Communication magazine series on IoT. His interests span from the field of AI/ML to Game theory.
M. Majid Butt [SM'15] is a Senior Specialist at Nokia Bell-Labs, France, and an adjunct Professor at Trinity College Dublin, Ireland. He has authored more than 70 peer-reviewed conference and journal articles and filed over 30 patents. He is IEEE Comsoc distinguished lecturer for the class 2022-23. He frequently gives invited and technical tutorial talks on various topics in IEEE conferences and serves as an associate editor for IEEE Communication Magazine, IEEE Open Journal of the Communication Society and IEEE Open Journal of Vehicular Technology.
István Z. Kovács [M’00] received his B.Sc. from "Politehnica" Technical University of Timişoara, Romania in 1989, his M.Sc.E.E. from École Nationale Supérieure des Télécommunications de Bretagne, France in 1996, and his Ph.D.E.E. in Wireless Communications from Aalborg University, Denmark in 2002. Currently he is senior research engineer at Nokia, Aalborg, Denmark, where he conducts research on machine learning-driven radio resource management and radio connectivity enhancements for non-terrestrial and aerial vehicle communications, in LTE and 5G networks.
Mohamad Assaad [SM'15] is a Professor at CentraleSupelec, France and a researcher at the Laboratory of Signals and Systems (CNRS). He has coauthored 1 book and more than 120 journal and conference papers and serves regularly as TPC cochair for top-tier international conferences. He is currently an Editor for the IEEE Wireless Communications Letters and Journal of Communications and Information Networks. His research interests include 5G and beyond systems, and Machine Learning in wireless networks.
Mohsen Guizani [F'09] is currently a Professor at the Machine Learning Department at the Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE. His main research interests are wireless communications and IoT security. He was elevated to the IEEE Fellow in 2009. He was listed as a Clarivate Analytics Highly Cited Researcher in Computer Science in 2019, 2020 and 2021. Dr. Guizani has won several research awards. He is the author of ten books and more than 800 publications.
tests/benchmark/datasets/annotations/cleaned/cleaned_ordinary_textbook_1d9a847603a5e37e379738316820850d.md 0 → 100644
View file @ 09117e54
# 数学新星问题征解
第十五期 (2016.06)
主持: 牟晓生
第一题. 设 $z_{1}, z_{2}, z_{3}$ 是单位复数. 证明存在单位复数 $z$ 使得:
$$
\frac{1}{\left|z-z_{1}\right|^{2}}+\frac{1}{\left|z-z_{2}\right|^{2}}+\frac{1}{\left|z-z_{3}\right|^{2}} \leq \frac{9}{4}
$$
(湖北武钢三中学生 王逸轩, 上海大学冷岗松 供题)
第二题. 如图, $D$ 是正三角形 $A B C$ 的边 $B C$ 上一点, $B D>C D$. 记 $O_{1}, I_{1}$ 为 $\triangle A B D$ 的外心与内心, $O_{2}, I_{2}$ 为 $\triangle A C D$ 的外心与内心. 圆 $I_{1}$ 与圆 $I_{2}$ 除 $B C$外的另一条外公切线交 $A B, A C$ 于 $P, Q$. 设直线 $P I_{1}$与 $Q I_{2}$ 交于 $R$, 而直线 $O_{1} I_{1}$ 与 $O_{2} I_{2}$ 交于 $T$. 证明: $A T^{2}=A R^{2}+A D \cdot B C$.
(广西钦州 卢圣 供题)
第三题. 给定正整数 $m, n$, 考虑在 $m \times n$ 白棋盘上先将一些格染成黑色. 在之后的每一时刻, 若存在一个白格至少与两个黑格相邻, 则可将它也染成黑色. 求最初至少要染多少个黑色格才能在某一时刻染黑整个棋盘?
(哈佛大学 牟晓生 供题)
第四题. $A B C$ 是一个三角形, 而 $P, Q, R$ 分别是 $B C, C A, A B$ 上的点。证明 $\triangle P Q R$ 的周长不小于 $\triangle A Q R, \triangle B R P, \triangle C P Q$ 周长的最小值.
(哈佛大学 牟晓生 供题)
tests/benchmark/datasets/annotations/cleaned/cleaned_research_report_1f978cd81fb7260c8f7644039ec2c054.md 0 → 100644
View file @ 09117e54
## 增持（维持）
所属行业：机械设备
当前价格(元): 82.42
## 证券分析师
倪正洋
资格编号：S0120521020003
邮箱: nizy@tebon.com.cn
## 研究助理
杨云道
邮箱: yangyx@tebon.com.cn
| 沪深 300 对比 | $1 \mathrm{M}$ | $2 \mathrm{M}$ | $3 \mathrm{M}$ |
| :--- | ---: | ---: | ---: |
| 绝对涨幅(\%) | 7.18 | 32.88 | 80.86 |
| 相对涨幅(\%) | 8.10 | 25.93 | 78.39 |
资料来源: 德邦研究所, 聚源数据
## 相关研究
1.《高测股份 (688556): 光伏金刚线及硅片切割代工业务推动公司 22Q1 业绩大超预期》, 2022.4.29
2.《光伏设备: 光伏高效电池扩产提速,关键设备商各领风骚》, 2022.4.10 3. 《高测股份 (688556.SH): 再签建湖 10GW 硅片切割代工产能，强化代工业务成长逻辑》, 2022.4.7
3.《高测股份 (688556.SH): 签订晶澳曲靖 2.2 亿元切割设备合同，看好 22 年代工业绩释放+HJT 切割工艺进步》, 2022.3.9
4.《高测股份 (688556.SH): 21 年业绩预告超市场预期，关注切片代工利润释放》, 2022.1.24
# 高测股份 $(688556.5 H):$ 扩产 4000 万公里金刚线，强化光伏碰片切割三元布局
## 投资要点
- 事件：公司拟与蓝关县人民政府签署的《壶关年产 12000 万千米金刚线项目投资协议书》，项目一期计划建设年产 4,000万千米金刚线产能，预计一期总投资额约 6.66 亿元; 后续年产 8,000 万千米金刚线项目尚未具体约定，存在较大不确定性。
- 顺应下游需求扩张, 金刚线产能快速扩产, 保证公司内供+外销。光伏金刚线需求 22 年提升源于两方面：1）2022 年光伏产业链景气度高涨，1-5 月光伏装机同比 $+24.4 \%$, 带动产业链各环节开工率提升, 硅片前期扩产产能逐步落地, 金刚线需求释放；2）由于多晶硅料价格持续维持高位，细线化、薄片化趋势加速，其中细线化要求金刚线线径由 40 线、 38 线向 36 线、 35 线进步, 带动单 GW 切割线耗不断提升。目前 36 线单 GW 切割线耗约 50 万公里, 较 38 线提升约 $30 \%$ 。公司于 2021 年对金刚线进行 “ 1 机 12 线” 技改，技改完成后，公司 22 年 1 季度产能 712 万公里, 年化产能超 2500 万公里。公司目前切片代工产能约 47GW, 对应远期金刚线产能超 2300 万公里。本次扩产再一次扩充公司金刚线产能, 强化金刚线产能内供+外销布局。
- 依托萦关低成本电价提升金刚线盈利能力, 顺应硅料节约持续推动细线化布局。公司在山西长治金刚线生产厂区采购电力的平均单价较青岛金刚线生产厂区采购电力的平均单价低, 2020 年度公司陆续将青岛的金刚线生产线搬迁到山西长治並关厂区，随着山西长治金刚线生产厂区金刚线产量增加，公司采购电力的平均单价呈下降趋势。目前公司电力采购单价从 2019 年 0.8 元/kwh 降低到 2022 年 Q1 的 0.39 元/kwh，並关后续拓展有望进一步降低公司金刚线电价成本。金刚线线径越细，锯㖓越小，切割时产生的锯㖓硅料损失越少，同样一根硅棒可切割加工出的硅片数量越多，制造硅片所需的硅材料越少。相同切割工艺下，金刚线越细，固结在钢线基体上的金刚石微粉颗粒越小，切割加工时对硅片的表面损伤越小，硅片表面质量越好，砝片 TTV 等质量指标表现也就越好。金刚线母线直径已由 2016 年的 80um 降至 2022 年上半年的 36、38、40um，此外高线速、柔性化和智能化等均是金刚线及切片技术进步方向, 公司在薄片、细线化、高线速、柔性智能化方面均有领先布局, 推动切割工艺持续进步。
- 切割工艺的持续进步领先, 是保障公司利润释放的核心壁垒。公司光伏硅片切割三元布局包括硅片切割及机加工设备、砝片切割耗材 (金刚线) 以及切割代工业务。公司 2021 年依托前期设备+耗材布局切割代工业务, 目前已公布 47GW 产能 (乐山5GW 示范基地、乐山 20GW 大硅片及配套项目、建湖一期 10GW 项目,建湖二期 $12 \mathrm{GW}$ 项目), 客户包括通威、京运通、美科及建湖周边电池企业。22 年底公司有望实现超 20GW 切割代工产能, 且当前终端客户主要为下游电池企业。客户选择切割代工模式的核心在于凭借高测的专业化服务实现快速上产, 同时可获得较自建硅片切割产能或购买硅片更多的超额利润。超额利润的核心在于高测股份的切割代工技术领先, 可实现更多的硅片切割红利, 并与客户共享。未来随着金刚线扩产和切割技术进步, 公司光伏硅片切割代工利润弹性有望持续释放。
- 盈利预测与投资建议：预计公司 2022-2024 年归母净利润 4.7、7.2、9.3 亿元,对应 PE 30、20、15 倍，维持 “增持” 评级。
- 风险提示：硅片扩产不及预期，公司代工业务利润波动风险，市场竞争加剧。
<table><thead><tr><th>股票数据</th><th></th></tr></thead><tr><td>总股本(百万股):</td><td>227.92</td></tr><tr><td>流通 A 股(百万股):</td><td>167.01</td></tr><tr><td>52 周内股价区间(元):</td><td>21.60-97.40</td></tr><tr><td>总市值(百万元):</td><td>18,785.44</td></tr><tr><td>总资产(百万元):</td><td>3,508.81</td></tr><tr><td>每股净资产(元):</td><td>5.50</td></tr><tr><td>咨料来源，公司公告</td><td></td></tr></table>
<table><thead><tr><th>主要财务数据及预测</th><th></th><th></th><th></th><th></th><th></th></tr></thead><tr><td></td><td>2020</td><td>2021</td><td>2022E</td><td>2023E</td><td>2024E</td></tr><tr><td>营业收入(百万元)</td><td>746</td><td>1,567</td><td>3,684</td><td>5,056</td><td>5,752</td></tr><tr><td>(+/-)YOY(%)</td><td>4.5\%</td><td>110.0\%</td><td>135.1\%</td><td>37.2\%</td><td>13.8\%</td></tr><tr><td>净利润(百万元)</td><td>59</td><td>173</td><td>471</td><td>717</td><td>933</td></tr><tr><td>(+/-)YOY(%)</td><td>83.8\%</td><td>193.4\%</td><td>172.8\%</td><td>52.2\%</td><td>30.1\%</td></tr><tr><td>全面摊薄 EPS(元)</td><td>0.43</td><td>1.07</td><td>2.91</td><td>4.43</td><td>5.77</td></tr><tr><td>毛利率(\%)</td><td>35.3\%</td><td>33.7\%</td><td>35.0\%</td><td>36.0\%</td><td>38.0\%</td></tr><tr><td>净资产收益率(\%)</td><td>6.0\%</td><td>15.0\%</td><td>27.9\%</td><td>28.8\%</td><td>26.5\%</td></tr></table>
资料来源: 公司年报 (2020-2021)，德邦研究所
备注: 净利润为归属母公司所有者的净利润
## 财务报表分析和预测
| 主要财务指标 | 2021 | $2022 E$ | $2023 E$ | $2024 E$ |
| :--- | ---: | ---: | ---: | ---: |
| 每股指标(元) | | | | |
| 每股收益 | 1.07 | 2.91 | 4.43 | 5.77 |
| 每股净资产 | 7.13 | 10.43 | 15.39 | 21.76 |
| 每股经营现金流 | 0.47 | 1.27 | 4.07 | 5.02 |
| 每股股利 | 0.11 | 0.11 | 0.11 | 0.11 |
| 价值评估(倍) | | | | |
| P/E | 82.90 | 30.47 | 20.02 | 15.38 |
| P/B | 12.44 | 8.50 | 5.76 | 4.08 |
| P/S | 8.52 | 3.62 | 2.64 | 2.32 |
| EV/EBITDA | 49.85 | 24.12 | 15.68 | 11.46 |
| 股息率\% | $0.1 \%$ | $0.1 \%$ | $0.1 \%$ | $0.1 \%$ |
| 盈利能力指标(\%) | | | | |
| 毛利率 | $33.7 \%$ | $35.0 \%$ | $36.0 \%$ | $38.0 \%$ |
| 净利润率 | $11.0 \%$ | $12.8 \%$ | $14.2 \%$ | $16.2 \%$ |
| 净资产收益率 | $15.0 \%$ | $27.9 \%$ | $28.8 \%$ | $26.5 \%$ |
| 资产回报率 | $5.3 \%$ | $7.9 \%$ | $8.5 \%$ | $9.2 \%$ |
| 投资回报率 | $15.3 \%$ | $25.9 \%$ | $24.6 \%$ | $23.7 \%$ |
| 盈利增长(\%) | | | | |
| 营业收入增长率 | $110.0 \%$ | $135.1 \%$ | $37.2 \%$ | $13.8 \%$ |
| EBIT 增长率 | $233.7 \%$ | $150.7 \%$ | $52.3 \%$ | $31.9 \%$ |
| 净利润增长率 | $193.4 \%$ | $172.8 \%$ | $52.2 \%$ | $30.1 \%$ |
| 偿倩能力指标 | | | | |
| 资产负债率 | $64.3 \%$ | $71.5 \%$ | $70.6 \%$ | $65.3 \%$ |
| 流动比率 | 1.2 | 1.2 | 1.3 | 1.4 |
| 速动比率 | 0.9 | 0.9 | 1.0 | 1.1 |
| 现金比率 | 0.2 | 0.1 | 0.2 | 0.3 |
| 经营效率指标 | | | | |
| 应收怅款周转天数 | 161.7 | 165.1 | 164.9 | 164.4 |
| 存货周转天数 | 196.1 | 170.0 | 180.0 | 190.0 |
| 总资产周转率 | 0.5 | 0.6 | 0.6 | 0.6 |
| 固定资产周转率 | 4.2 | 8.6 | 10.3 | 11.1 |
| 现金流量表(百万元) | 2021 | $2022 E$ | 2023E | 2024E |
| :--- | ---: | ---: | ---: | ---: |
| 净利润 | 173 | 471 | 717 | 933 |
| 少数股东损益 | 0 | 0 | 0 | 0 |
| 非现金支出 | 107 | 114 | 133 | 147 |
| 非经营收益 | 17 | 1 | 4 | 14 |
| 营运资金变动 | -220 | -382 | -195 | -283 |
| 经营活动现金流 | 76 | 205 | 658 | 812 |
| 资产 | -83 | -184 | -203 | -169 |
| 投资 | 229 | 0 | 0 | 0 |
| 其他 | 6 | 9 | 13 | 14 |
| 投资活动现金流 | 151 | -175 | -190 | -155 |
| 债权募资 | -80 | 39 | 321 | 64 |
| 股权募资 | 0 | 0 | 0 | 0 |
| 其他活 | -21 | -3 | -14 | -25 |
| 融资活动现金流 | -101 | 36 | 307 | 39 |
| 现金净流量 | 127 | 66 | 775 | 696 |
备注: 表中计算估值指标的收盘价日期为 7 月 19 日
资料来源: 公司年报 (2020-2021), 德邦研究所
| 利润表(百万元) | 2021 | 2022E | 2023E | 2024E |
| :---: | :---: | :---: | :---: | :---: |
| 营业总收入 | 1,567 | 3,684 | 5,056 | 5,752 |
| 营业成本 | 1,038 | 2,394 | 3,236 | 3,567 |
| 毛利率\% | $33.7 \%$ | $35.0 \%$ | $36.0 \%$ | $38.0 \%$ |
| 营业税金及附加 | 6 | 18 | 25 | 29 |
| 营业税金率\% | $0.4 \%$ | $0.5 \%$ | $0.5 \%$ | $0.5 \%$ |
| 营业费用 | 63 | 147 | 193 | 209 |
| 营业费用率\% | $4.0 \%$ | $4.0 \%$ | $3.8 \%$ | $3.6 \%$ |
| 管理费用 | 131 | 313 | 409 | 444 |
| 管理费用率\% | $8.4 \%$ | $8.5 \%$ | $8.1 \%$ | $7.7 \%$ |
| 研发费用 | 117 | 276 | 379 | 431 |
| 研发费用率\% | $7.5 \%$ | $7.5 \%$ | $7.5 \%$ | $7.5 \%$ |
| EBIT | 213 | 534 | 814 | 1,074 |
| 财务费用 | 7 | 1 | 11 | 19 |
| 财务费用率\% | $0.4 \%$ | $0.0 \%$ | $0.2 \%$ | $0.3 \%$ |
| 资产减值损失 | -33 | -63 | -86 | -98 |
| 投资收益 | 5 | 9 | 13 | 14 |
| 营业利润 | 212 | 531 | 800 | 1,040 |
| 营业外收支 | -25 | -8 | -3 | -3 |
| 利润总额 | 187 | 523 | 797 | 1,037 |
| EBITDA | 282 | 582 | 865 | 1,129 |
| 所得税 | 14 | 52 | 80 | 104 |
| 有效所得税率\% | $7.7 \%$ | $10.0 \%$ | $10.0 \%$ | $10.0 \%$ |
| 少数股东损益 | 0 | 0 | 0 | $\mathbf{0}-1-2$ |
| 归属母公司所有者净利润 | 173 | 471 | 717 | 933 |
| 资产负债表(百万元) | 2021 | 2022E | 2023E | $2024 E$ |
| :---: | :---: | :---: | :---: | :---: |
| 货币资金 | 427 | 494 | 1,269 | 1,965 |
| 应收账款及应收票据 | 1,173 | 2,806 | 3,798 | 4,344 |
| 存货 | 558 | 1,115 | 1,596 | 1,857 |
| 其它流动资产 | 266 | 578 | 736 | 778 |
| 流动资产合计 | 2,424 | 4,992 | 7,400 | 8,943 |
| 长期股权投资 | 0 | 0 | 0 | 0 |
| 固定资产 | 370 | 429 | 491 | 516 |
| 在建工程 | 169 | 183 | 205 | 226 |
| 无形资产 | 42 | 56 | 69 | 80 |
| 非流动资产合计 | 811 | 940 | 1,087 | 1,198 |
| 资产总计 | 3,235 | 5,932 | 8,487 | 10,141 |
| 短期借款 | 28 | 68 | 388 | 452 |
| 应付票据及应付账款 | 1,401 | 3,197 | 4,302 | 4,760 |
| 预收账款 | 0 | 0 | 0 | 0 |
| 其它流动负债 | 560 | 887 | 1,214 | 1,314 |
| 流动负债合计 | 1,989 | 4,152 | 5,904 | 6,527 |
| 长期借款 | 0 | 0 | 0 | 0 |
| 其它长期负债 | 92 | 92 | 92 | 92 |
| 非流动负债合计 | 92 | 92 | 92 | 92 |
| 负债总计 | 2,081 | 4,243 | 5,996 | 6,619 |
| 实收资本 | 162 | 162 | 162 | 162 |
| 普通股股东权益 | 1,154 | 1,688 | 2,491 | 3,522 |
| 少数股东权益 | 0 | 0 | 0 | 0 |
| 负债和所有者权益合计 | 3,235 | 5,932 | 8,487 | 10,141 |
## 信息披露
## 分析师与研究助理简介
倪正洋，2021 年加入德邦证券，任研究所大制造组组长、机械行业首席分析师，拥有 5 年机械研究经验，1 年高端装备产业经验，南京大学材料学学士、上海交通大学材料学硕士。2020 年获得 iFinD 机械行业最具人气分析师, 所在团队曾获机械行业 2019 年新财富第三名，2017 年新财富第二名，2017 年金牛奖第二名，2016 年新财富第四名。
## 分析师声明
本人具有中国证券业协会授予的证券投资咨询执业资格，以勤勉的职业态度，独立、客观地出具本报告。本报告所采用的数据和信息均来自市场公开信息, 本人不保证该等信息的准确性或完整性。分析逻辑基于作者的职业理解，清晰准确地反映了作者的研究观点，结论不受任何第三方的授意或影响，特此声明。
## 投资评级说明
1.投资评级的比较和评级标准:
以报告发布后的 6 个月内的市场表现为比较标准，报告发布日后 6 个月内的公司股价（或行业指数）的张跌幅相对同期市场基准指数的涨跌幅;
2.市场基准指数的比较标准:
A 股市场以上证综指或深证成指为基准；香港市场以恒生指数为基准；美国市场以标普 500 或纳斯达克综合指数为基准。
<table>
<tr>
<td rowspan="11">1. 投资评级的比较和评级标准: 以报告发布后的 6 个月内的市场表 现为比较标准，报告发布日后 6 个 月内的公司股价(或行业指数)的 涨跌幅相对同期市场基准指数的涨 跌幅：<br> 2. 市场基准指数的比较标准: A股市场以上证综指或深证成指为基 准; 香港市场以恒生指数为基准; 美 国市场以标普500或纳斯达克综合指 数为基准。</td>
</tr>
<tr>
<td>类型</td>
<td>评级</td>
<td>说明</td>
</tr>
<td rowspan="5">股票评级</td>
</tr>
<tr>
<td>买入</td>
<td>相对强于市场表现 20%以上;</td>
</tr>
<tr>
<td>增持</td>
<td>相对强于市场表现 5% 20%;</td>
</tr>
<tr>
<td>中性</td>
<td>相对市场表现在-5% +5%之间波动；</td>
</tr>
<tr>
<td>减持</td>
<td>相对弱于市场表现 5%以下。</td>
</tr>
<tr>
<td rowspan="4">行业投资评级</td>
</tr>
<tr>
<td>优于大市</td>
<td>预期行业整体回报高于基准指数整体水平10%以上;</td>
</tr>
<tr>
<td>中性</td>
<td>预期行业整体回报介于基准指数整体水平-10%与 10%之间;</td>
</tr>
<tr>
<td>弱于大市</td>
<td>预期行业整体回报低于基准指数整体水平 10%以下。</td>
</tr>
<tr>
</table>
## 法律声明
本报告仅供德邦证券股份有限公司（以下简称 “本公司”）的客户使用。本公司不会因接收人收到本报告而视其为客户。在任何情况下，本报告中的信息或所表述的意见并不构成对任何人的投资建议。在任何情况下，本公司不对任何人因使用本报告中的任何内容所引致的任何损失负任何责任。
本报告所载的资料、意见及推测仅反映本公司于发布本报告当日的判断，本报告所指的证券或投资标的的价格、价值及投资收入可能会波动。在不同时期，本公司可发出与本报告所载资料、意见及推测不一致的报告。
市场有风险，投资需谨慎。本报告所载的信息、材料及结论只提供特定客户作参考，不构成投资建议，也没有考虑到个别客户特殊的投资目标、财务状况或需要。客户应考虑本报告中的任何意见或建议是否符合其特定状况。在法律许可的情况下，德邦证券及其所属关联机构可能会持有报告中提到的公司所发行的证券并进行交易，还可能为这些公司提供投资银行服务或其他服务。
本报告仅向特定客户传送，未经德邦证券研究所书面授权，本研究报告的任何部分均不得以任何方式制作任何形式的拷贝、复印件或复制品，或再次分发给任何其他人，或以任何侵犯本公司版权的其他方式使用。所有本报告中使用的商标、服务标记及标记均为本公司的商标、服务标记及标记。如欲引用或转载本文内容, 务必联络德邦证券研究所并获得许可, 并需注明出处为德邦证券研究所，且不得对本文进行有悖原意的引用和删改。
根据中国证监会核发的经营证券业务许可，德邦证券股份有限公司的经营范围包括证券投资咨询业务。
\ No newline at end of file
tests/benchmark/env.sh 0 → 100644
View file @ 09117e54
conda create -n MinerU python=3.10
conda activate MinerU
pip install magic-pdf
pip install magic-pdf[full-cpu]
pip install detectron2 --extra-index-url https://myhloli.github.io/wheels/
git lfs install
git lfs clone https://huggingface.co/wanderkid/PDF-Extract-Kit
#cp magic-pdf.template.json ~/magic-pdf.json
\ No newline at end of file
tests/benchmark/magic-pdf.json 0 → 100644
View file @ 09117e54
{
"bucket_info":{
"bucket-name-1":["ak", "sk", "endpoint"],
"bucket-name-2":["ak", "sk", "endpoint"]
},
"temp-output-dir":"/tmp",
"models-dir":"/tmp/models",
"device-mode":"cpu"
}
\ No newline at end of file
tests/benchmark/pre_clean.py 0 → 100644
View file @ 09117e54
"""
clean data
"""
import argparse
import os
import re
import htmltabletomd # type: ignore
import pypandoc
import argparse
parser = argparse.ArgumentParser(description="get tool type")
parser.add_argument(
"--tool_name",
type=str,
required=True,
help="input tool name",
)
parser.add_argument(
"--download_dir",
type=str,
required=True,
help="input download dir",
)
args = parser.parse_args()
def clean_markdown_images(content):
"""
clean markdown images
"""
pattern = re.compile(r'!\[[^\]]*\]\([^)]*\)', re.IGNORECASE)
cleaned_content = pattern.sub('', content)
return cleaned_content
def clean_ocrmath_photo(content):
"""
clean ocrmath photo
"""
pattern = re.compile(r'\\includegraphics\[.*?\]\{.*?\}', re.IGNORECASE)
cleaned_content = pattern.sub('', content)
return cleaned_content
def convert_html_table_to_md(html_table):
"""
convert html table to markdown table
"""
lines = html_table.strip().split('\n')
md_table = ''
if lines and '<tr>' in lines[0]:
in_thead = True
for line in lines:
if '<th>' in line:
cells = re.findall(r'<th>(.*?)</th>', line)
md_table += '| ' + ' | '.join(cells) + ' |\n'
in_thead = False
elif '<td>' in line and not in_thead:
cells = re.findall(r'<td>(.*?)</td>', line)
md_table += '| ' + ' | '.join(cells) + ' |\n'
md_table = md_table.rstrip() + '\n'
return md_table
def convert_latext_to_md(content):
"""
convert latex table to markdown table
"""
tables = re.findall(r'\\begin\{tabular\}(.*?)\\end\{tabular\}', content, re.DOTALL)
placeholders = []
for table in tables:
placeholder = f"<!-- TABLE_PLACEHOLDER_{len(placeholders)} -->"
replace_str = f"\\begin{{tabular}}{table}cl\\end{{tabular}}"
content = content.replace(replace_str, placeholder)
try:
pypandoc.convert_text(replace_str, format="latex", to="md", outputfile="output.md", encoding="utf-8")
except:
markdown_string = replace_str
else:
markdown_string = open('output.md', 'r', encoding='utf-8').read()
placeholders.append((placeholder, markdown_string))
new_content = content
for placeholder, md_table in placeholders:
new_content = new_content.replace(placeholder, md_table)
# 写入文件
return new_content
def convert_htmltale_to_md(content):
"""
convert html table to markdown table
"""
tables = re.findall(r'<table>(.*?)</table>', content, re.DOTALL)
placeholders = []
for table in tables:
placeholder = f"<!-- TABLE_PLACEHOLDER_{len(placeholders)} -->"
content = content.replace(f"<table>{table}</table>", placeholder)
try:
convert_table = htmltabletomd.convert_table(table)
except:
convert_table = table
placeholders.append((placeholder,convert_table))
new_content = content
for placeholder, md_table in placeholders:
new_content = new_content.replace(placeholder, md_table)
# 写入文件
return new_content
def clean_data(prod_type, download_dir):
"""
clean data
"""
tgt_dir = os.path.join(download_dir, prod_type, "cleaned")
if not os.path.exists(tgt_dir):
os.makedirs(tgt_dir)
source_dir = os.path.join(download_dir, prod_type)
filenames = os.listdir(source_dir)
for filename in filenames:
if filename.endswith('.md'):
input_file = os.path.join(source_dir, filename)
output_file = os.path.join(tgt_dir, "cleaned_" + filename)
with open(input_file, 'r', encoding='utf-8') as fr:
content = fr.read()
new_content = clean_markdown_images(content)
new_content = convert_html_table_to_md(new_content)
new_content = convert_latext_to_md(new_content)
new_content = convert_htmltale_to_md(new_content)
with open(output_file, 'w', encoding='utf-8') as fw:
fw.write(new_content)
if __name__ == '__main__':
tool_type = args.tool_name
download_dir = args.download_dir
clean_data(tool_type, download_dir)
tests/benchmark/result.json 0 → 100644
View file @ 09117e54
{"average_sim_score":0, "average_edit_distance":0, "average_bleu_score": 0}
\ No newline at end of file
tests/benchmark/scoring.py 0 → 100644
View file @ 09117e54
import math
from rapidfuzz import fuzz
import re
import regex
from statistics import mean
CHUNK_MIN_CHARS = 25
def chunk_text(text, chunk_len=500):
chunks = [text[i:i+chunk_len] for i in range(0, len(text), chunk_len)]
chunks = [c for c in chunks if c.strip() and len(c) > CHUNK_MIN_CHARS]
return chunks
def overlap_score(hypothesis_chunks, reference_chunks):
if len(reference_chunks) > 0:
length_modifier = len(hypothesis_chunks) / len(reference_chunks)
else:
length_modifier = 0
search_distance = max(len(reference_chunks) // 5, 10)
chunk_scores = []
for i, hyp_chunk in enumerate(hypothesis_chunks):
max_score = 0
total_len = 0
i_offset = int(i * length_modifier)
chunk_range = range(max(0, i_offset-search_distance), min(len(reference_chunks), i_offset+search_distance))
for j in chunk_range:
ref_chunk = reference_chunks[j]
score = fuzz.ratio(hyp_chunk, ref_chunk, score_cutoff=30) / 100
if score > max_score:
max_score = score
total_len = len(ref_chunk)
chunk_scores.append(max_score)
return chunk_scores
def score_text(hypothesis, reference):
# Returns a 0-1 alignment score
hypothesis_chunks = chunk_text(hypothesis)
reference_chunks = chunk_text(reference)
chunk_scores = overlap_score(hypothesis_chunks, reference_chunks)
if len(chunk_scores) > 0:
mean_score = mean(chunk_scores)
return mean_score
else:
return 0
#return mean(chunk_scores)
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment