DeePMD-kit

DeePMD-kit documentation: https://deepmd.readthedocs.io

1. Introduction

DeePMD-kit is a machine learning-based tool that fits first-principles (DFT) potential energy surfaces (PES) to be used in molecular dynamics (MD) simulations. It provides ab initio accuracy at a fraction of the computational cost and integrates with LAMMPS, GROMACS, OpenMM, etc.

2. Workflow Overview

Prepare Data: Convert AIMD/DFT data into DeePMD-kit format.
Train: Use dp train input.json to fit a model.
Freeze: Convert trained models into .pb files.
Compress (Optional): Optimize .pb for efficiency.
Test: Validate energy/force predictions.
Run MD: Use the trained model in LAMMPS via pair_style deepmd.

3. Practical Guide

3.1. Data Preparation

DeePMD-kit uses a compressed NumPy-based format for efficient storage of atomic structures and properties. To convert raw simulation outputs into this format, we rely on the dpdata tool, which supports VASP, CP2K, Gaussian, Quantum Espresso, ABACUS, and LAMMPS.

Convert first-principles data to DeePMD-kit format:

import dpdata, numpy as np

data = dpdata.LabeledSystem("00.data/abacus_md", fmt="abacus/md")
index_val = np.random.choice(len(data), 40, replace=False)
index_train = list(set(range(len(data))) - set(index_val))

data.sub_system(index_train).to_deepmd_npy("00.data/training_data")
data.sub_system(index_val).to_deepmd_npy("00.data/validation_data")

To convert other formats, change the fmt argument to one of:

"vasp/poscar"
"cp2k/xyz"
"gaussian/log"
"qe/sssp"
"lammps/dump"

3.2. Training Configuration

Define input.json for training:

{
  "model": {
    "type_map": ["H", "C"],
    "descriptor": { "type": "se_e2_a", "rcut": 6.0, "neuron": [25, 50, 100] },
    "fitting_net": { "neuron": [240, 240, 240] }
  },
  "training": {
    "training_data": { "systems": ["../00.data/training_data"] },
    "validation_data": { "systems": ["../00.data/validation_data"] },
    "numb_steps": 10000
  }
}

3.3. Train the Model

Run:

dp train input.json

Monitor training loss in lcurve.out.

3.4. Freeze & Optimize Model

Convert to a frozen model:

dp freeze -o graph.pb

(Optional) Compress the model:

dp compress -i graph.pb -o compress.pb

3.5. Model Testing

Evaluate performance:

dp test -m graph.pb -s ../00.data/validation_data

For visualization:

import dpdata, matplotlib.pyplot as plt

val_system = dpdata.LabeledSystem("../00.data/validation_data", fmt="deepmd/npy")
prediction = val_system.predict("graph.pb")

plt.scatter(val_system["energies"], prediction["energies"], alpha=0.5)
plt.xlabel("DFT Energy (eV)")
plt.ylabel("DP Predicted Energy (eV)")
plt.show()

3.6. Running MD with LAMMPS

Write in.lammps:

units metal
atom_style atomic
read_data conf.lmp
pair_style deepmd graph.pb
pair_coeff * *
timestep 0.001
run 5000