DeepMD 使用

介绍

DeePMD 使用教程、科研案例、问题收集合集 - Bohrium

使用深度神经网络构建势函数

版本：初始版本（v0）、v1、v2

DP library：DP 模型数据库（类似 Material Project）

exploration、labeling、train 三个过程

learning/notes/DeePMD/DeePMD-kit安装.md at main · wangchr1617/learning · GitHub

learning/notes/DeePMD/DeePMD-kit入门.md at main · wangchr1617/learning · GitHub

DeepMD ASE calculator

1
2
3
4
5
6
7
8
9
from deepmd.calculator import DP

calc = DP(model=...)

atoms.calc = calc

energy = atoms.get_potential_energy()
forces = atoms.get_forces()
stress = atoms.get_stress()

描述符

1
2
3
4
5
sel_a

se_e2_a
se_e3
se_atten

模型压缩

实例

安装

DeePMD-kit

含 CPU 和 GPU 版本

1

dpgen

1
2
3
4
5
6
7
8
9
10
# PyPI 安装
pip install dpgen
# Conda 安装
conda install -c conda-forge dpgen
# 源码安装
git clone https://github.com/deepmodeling/dpgen && pip install ./dpgen


# dpdata
pip install -U dpdata

使用

工具

机器学习力场训练集可视化 - 知乎

1
2
3
4
5
6
7
8
from deepmd.infer import DeepPot

dp = DeepPot(args.model)

# 获取 DP 描述符
descriptor = dp.eval_descriptor()

dp.eval()

DeePMD-kit

训练

image.png

训练数据中必要的信息：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
# 训练集数据结构
training_data
├── set.000
│   ├── box.npy
│   ├── coord.npy
│   ├── energy.npy
│   ├── force.npy
│   └── virial.npy
├── type.raw
└── type_map.raw


coord.npy      # 原子坐标
type.raw       # 元素种类
energy.npy     # 体系能量
force.npy      # 原子受力
box.npy        # 模拟盒子，若为非周期性体系，请在训练文件里准备一个超大周期边界条件

virial.raw     # 维里应力
  • 模型训练参数
1
2
3
4
model
learning_rate
loss
training
  • 示例
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
"model":{
    "type_map":    ["H", "C"],                 
    "descriptor":{
        "type":            "se_e2_a",          
        "rcut":            6.00,               
        "rcut_smth":       0.50,               
        "sel":             "auto",             
        "neuron":          [25, 50, 100],       
        "resnet_dt":       false,
        "axis_neuron":     16,                  
        "seed":            1,
        "_comment":        "that's all"
        },
    "fitting_net":{
        "neuron":          [240, 240, 240],    
        "resnet_dt":       true,
        "seed":            1,
        "_comment":        "that's all"
    },
    "_comment":    "that's all"'
},
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# 开始训练
dp train input.json
# 重启训练
dp train input.json --restart model.ckpt

train.log           # 训练记录文件
lcurve.out          # 学习曲线



# 提取经过训练的神经网络模型，生成冻结模型 graph.pb
dp freeze -o graph.pb
# 压缩
dp compress -i graph.pb -o graph-compress.pb

# 检查势函数模型
dp test -m graph.pb -s ...

dpgen

构建训练集、训练

同步学习

image.png

autoset 功能

使用 DeePMD-kit 同时训练参数初始化不同的 4 个势函数

1
2
3
4
5
6
7
8
9
10
11
12
13

# dpgen 命令
init_bulk
init_surf
run

dpgen sub-command PARAM MACHINE
# 参数
PARAM              # 输入参数文件
MACHINE            # 机器配置文件


dpgen.log           #

mdapy 库可生成微扰结构

1
2
3
4
5
import mdapy as mp

mp.init()

pert = mp.PerturbModel()

DPMD - Deep Potential Molecular Dynamics-Prepare Dataset

机器学习势预测精度用 MAE 评估

Deepmd-kit & DPGEN 使用笔记

离线安装

给离线安装的 conda 环境添加环境名

1
ln -s ~/src/deepmd-kit ~/src/miniconda3/envs/<new_env_name>
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
deepmd-kit will now be installed into this location:
/home/yangsl/deepmd-kit

  - Press ENTER to confirm the location
  - Press CTRL-C to abort the installation
  - Or specify a different location below

[/home/yangsl/deepmd-kit] >>> $HOME/src/deepmd-kit
PREFIX=/home/yangsl/src/deepmd-kit
Unpacking payload ...
Notes:
The off-line packages and conda packages require the GNU C Library 2.17 or above[1]. The GPU version requires compatible NVIDIA driver to be installed in advance[2]. It is possible to force conda to override detection when installation[3] (such as CONDA_OVERRIDE_CUDA), but these requirements are still necessary during runtime.

[1] The GNU C Library. https://www.gnu.org/software/libc/
[2] Minor Version Compatibility. NVIDIA Data Center GPU Driver Documentation. https://docs.nvidia.com/deploy/cuda-compatibility/index.html#minor-version-compatibility
[3] Overriding detected packages. conda documentation. https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-virtual.html#overriding-detected-packages


Installing base environment...


Downloading and Extracting Packages


Downloading and Extracting Packages

Preparing transaction: done
Executing transaction: done
Please activate the environment before using the packages:

source /path/to/deepmd-kit/bin/activate /path/to/deepmd-kit

The following executable files have been installed:
1. DeePMD-kit CLi: dp -h
2. LAMMPS: lmp -h
3. DeePMD-kit i-Pi interface: dp_ipi
4. MPICH: mpirun -h
5. Horovod: horovod -h

The following Python libraries have been installed:
1. deepmd
2. dpdata
3. pylammps

If you have any questions, seek help from https://github.com/deepmodeling/deepmd-kit/discussions

这两个 raw 文件内容对训练是否会有影响？

通配符获取 outcar deepmd 数据，能量顺序不是按第 1-4 步来的

默认 240

将 ICET 的 data 转换成 deepmd 格式

water se_e2_a 80*3 训练集 80 测试集 type.new 192 行

1
2
3
4


disp_freq

dpdata

  • dpdata：将多种构型文件格式转换成 deepmd 格式（转换成其他格式的功能一般）
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
import dpdata

d_outcar = dpdata.LabeledSystem("OUTCAR", fmt="vasp/outcar")

d_outcar["energies"]
d_outcar["forces"]
d_outcar["virials"]

d_outcar.get_nframes()
d_outcar.get_natoms()

# keys；无 stress key
atom_names
atom_numbs
atoms_types
cells                    # 获取的是 VOLUME and BASIS
coords
energies                 # 获取的是 free  energy   TOTEN 
forces
virials                  # 获取的是 FORCE on cell =-STRESS in cart. coord.  units 中的 Total


dpdata.LabeledSystem('OUTCAR').to()


# 常用格式
"deepmd/raw"
"deepmd/npy"
"vasp/poscar"
"vasp/contcar"
"vasp/string"
"vasp/outcar"
"lammps/lmp"
"lammps/dump"
"pymatgen/structure"
"pymatgen/molecule" 
"pymatgen/computedstructureentry"
"ase/structure"                  


# 预测数据和原始数据之间的比较
training_systems = dpdata.LabeledSystem()
predict = training_systems.predict("graph.pb")

training_systems["energies"]
predict["energies"]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
import dpdata

pert_num = 2
perturbed_system = dpdata.System("POSCAR").perturb(
    pert_num=pert_num,
    cell_pert_fraction=0.03,
    atom_pert_distance=0.01,
    atom_pert_style="uniform",
    atom_pert_prob=1.0,
)

# 保存结构
for i in range(pert_num):
    # 方式 1
    perturbed_system.to(
        "vasp/poscar",
        f"pert/POSCAR.{i}",
        frame_idx=i,
    )
    # 方式 2
    perturbed_system.sub_system(i).to(
        "vasp/poscar",
        f"pert/POSCAR.{i}",
    )

# 方式 3
atoms_list = perturbed_system.to("ase/structure")
write(xyz_fn, atoms_list, format="extxyz", append=True)


# 会调用 get_cell_perturb_matrix()、get_atom_perturb_vector() 函数


# 参数
pert_num                     # 每帧构型生成扰动新构型的数目
cell_pert_fraction           # cell 变形比例；对角线 [-x, +x] 均匀分布扰动，非对角线 [-0.5x, +0.5x] 均匀分布扰动
atom_pert_distance           # 原子移/扰动距离
atom_pert_style              # 原子移/扰动距离的分布；normal uniform const
atom_pert_prob               # 每帧构型中要扰动的原子数目比例

势函数使用

1.4. Install LAMMPS — DeePMD-kit documentation

1.2. Install from source code — DeePMD-kit documentation

DeePMD 与 LAMMPS 的接口

先安装 DeePMD-kit 的 C++ 接口

1
2
3
4
5
cd $deepmd_source_dir/source
mkdir build
cd build

cmake -DUSE_TF_PYTHON_LIBS=TRUE -DCMAKE_INSTALL_PREFIX=$deepmd_root ..
1
2
3
4
5
6
7
8
9
cd $deepmd_source_dir/source/build
make lammps          # 生成 USER-DEEPMD module


cp -r $deepmd_source_dir/source/build/USER-DEEPMD path_lammps/src

# 安装 deempd package
make yes-extra-fix
make yes-user-deepmd
1
2
pair_style  deepmd graph.pb
pair_coeff  * *