# DeePKS-kit(Deep Kohn-Sham)

# 模型介绍
DeePKS-kit是一个为量子化学系统生成精确能量泛函的程序，DeePKS通过机器学习对于低精度的DFT泛函进行优化，利用神经网络修正项去学习基线泛函（低精度、低成本）与目标第一性原理方法（高精度、高成本）计算得出的能量与力的差值。

# 模型结构

DeePKS-kit一端与PyTorch接口，另一端与PySCF接口，PySCF是一个从头算计算化学程序，为量子化学代码开发和计算提供了一个简单、轻量级、高效的平台。DeePKS-kit支持作者之前开发的DeePHF和DeePKS方法。此外，它还提供了一定的灵活性，例如，修改模型结构，改变训练方案，接口其他量子化学包等。DeePKS-kit由三个主要模块组成，处理以下任务:(1)使用预先计算的描述符和标签训练(微扰)神经网络(NN)能量函数;(2)利用所提供的能量泛函求解给定系统的自洽场方程;(3)通过迭代调用任务(1)和(2)来学习自一致的能量函数。

![](./assets/deepks-kit.png)

DeePKS-kit架构示意图及工作流程如上图，上图说明了整个迭代学习过程的主要步骤。其中，左下是神经网络(NN)能量函数的训练。描述符从给定的分子轨道计算，并用作神经网络模型的输入。随机梯度下降(SGD)训练是使用PyTorch库实现的。右下是求解广义Kohn-Sham自洽场(SCF)方程。从训练好的神经网络泛函中计算XC电位。求解器是作为PySCF库的一个新类实现的。DeePKS-kit还提供了一个用户友好的界面，将上述模块与一些辅助功能组合成一个命令行工具。其中包含五个子命令： 
- `train`: 训练基于神经网络的后高频能量泛函模型
- `test`: 使用给定数据测试后高频模型并显示统计数据
- `scf`: 使用给定的能量模型运行自洽场计算
- `stats`: 收集和打印云函数的统计结果
- `iterate`: 通过组合上面的四个命令来迭代训练自洽模型

## 环境配置


## 性能和准确率数据

DCU测试平台：Z100
NVIDIA测试平台：A800

**1)收敛性测试**
Deepks的输出结果文件为test.out文件，默认运行的epoch数为10000，从左到右输出的指标分别是训练步骤、训练损失、验证损失、学习率、训练时间、验证时间。其中截取了最后的epoch部分的数值，其中看DCU和A800最后的结果基本一致。
[DCU文件信息]
![](./assets/deepks-1.png)
![](./assets/deepks-2.png)
![](./assets/deepks-3.png)
![](./assets/deepks-4.png)
[A800的文件输出信息]
![](./assets/deepks-5.png)
![](./assets/deepks-6.png)
![](./assets/deepks-7.png)
![](./assets/deepks-8.png)

整体收敛趋势如下图所示，依次为DCU和GPU。
![](./assets/deepks-9.png)
![](./assets/deepks-10.png)

预测值与实际值对照如下，依次为DCU和GPU。
![](./assets/deepks-11.png)
![](./assets/deepks-12.png)

**2)性能测试**
具体性能对比如下表：
| DCU | GPU |
| ------ | ------ |
| cell | cell |
| cell | cell |


## 源码仓库及问题反馈

DeePKS-kit is a pure python library so it can be installed following the standard `git clone` then `pip install` procedure. Note that the two main requirements `pytorch` and `pyscf` will not be installed automatically so you will need to install them manually in advance. Below is a more detailed instruction that includes installing the required libraries in the environment.

We use `conda` here as an example. So first you may need to install [Anaconda](https://docs.anaconda.com/anaconda/install/) or [Miniconda](https://docs.conda.io/en/latest/miniconda.html).

To reduce the possibility of library conflicts, we suggest create a new environment (named `deepks`) with basic dependencies installed (optional):
```bash
conda create -n deepks numpy scipy h5py ruamel.yaml paramiko
conda activate deepks
```
Now you are in the new environment called `deepks`.
Next, install [PyTorch](https://pytorch.org/get-started/locally/) 
```bash
# assuming a GPU with cudatoolkit 10.2 support
conda install pytorch cudatoolkit=10.2 -c pytorch
```
and [PySCF](https://github.com/pyscf/pyscf).
```bash
# the conda package does not support python >= 3.8 so we use pip
pip install pyscf
```

Once the environment has been setup properly, using pip to install DeePKS-kit:
```bash
pip install git+https://github.com/deepmodeling/deepks-kit/
```

## Usage

An relatively detailed decrisption of the `deepks-kit` library can be found in [here](https://arxiv.org/pdf/2012.14615.pdf). Please also refer to the reference for the description of methods.

Please see [`examples`](./examples) folder for the usage of `deepks-kit` library. A detailed example with executable data for single water molecules can be found [here](./examples/water_single). A more complicated one for training water clusters can be found [here](./examples/water_cluster).

Check [this input file](./examples/water_cluster/args.yaml) for detailed explanation for possible input parameters, and also [this one](./examples/water_cluster/shell.yaml) if you would like to run on local machine instead of using Slurm scheduler.

## References

[1] Chen, Y., Zhang, L., Wang, H. and E, W., 2020. Ground State Energy Functional with Hartree–Fock Efficiency and Chemical Accuracy. The Journal of Physical Chemistry A, 124(35), pp.7155-7165.

[2] Chen, Y., Zhang, L., Wang, H. and E, W., 2021. DeePKS: A Comprehensive Data-Driven Approach toward Chemically Accurate Density Functional Theory. Journal of Chemical Theory and Computation, 17(1), pp.170–181.