LAMMPS + DeePMD Potential

When to Use

• User wants to run MD with a trained DeePMD model
• User needs large-scale MD (10K-1M atoms) at near-DFT accuracy
• User wants to study diffusion, phase transitions, or surface reactions with ML potential
• User has a frozen .pb model file

Prerequisites

1. LAMMPS compiled with DEEPMD package (lmp -h | grep DEEPMD)
2. Frozen DeePMD model file (.pb)
3. Initial structure (LAMMPS data file or from CatGo viewer)
4. Know the type_map used during model training

Workflow Steps

1. Verify structure

catgo_view(action="get_state")

2. Create workflow

catgo_workflow_engine(action="create", params={"name": "LAMMPS DeePMD NVT 300K"})

3. Add LAMMPS task

catgo_workflow_engine(action="add_task", params={
  "workflow_id": "wf_xxx",
  "task_type": "shell",
  "name": "lmp_dpmd",
  "command": "lmp -in lammps.in > lammps.log 2>&1",
  "input_files": {
    "lammps.in": "<input script>",
    "frozen_model.pb": "{{model_path}}"
  },
  "system_name": "TiO2_md"
})

LAMMPS Input Template — NVT

units           metal
boundary        p p p
atom_style      atomic

read_data       structure.lmp

pair_style      deepmd frozen_model.pb
pair_coeff      * *

neighbor        2.0 bin
neigh_modify    every 1 delay 0 check yes

# Velocities
velocity        all create 300.0 12345 dist gaussian

# NVT thermostat
fix             1 all nvt temp 300.0 300.0 0.1

# Timestep (ps in metal units)
timestep        0.001

# Output
thermo          100
thermo_style    custom step temp pe ke etotal press vol

dump            1 all custom 100 traj.lammpstrj id type x y z fx fy fz
dump_modify     1 sort id

# Restart
restart         10000 restart.*.data

run             100000

LAMMPS Input Template — NPT

Replace the fix line:

fix             1 all npt temp 300.0 300.0 0.1 iso 0.0 0.0 1.0

Model Deviation (Multi-Model)

For active learning or reliability checking, use multiple models:

pair_style      deepmd model_0.pb model_1.pb model_2.pb model_3.pb out_freq 100 out_file model_devi.out
pair_coeff      * *

This writes model_devi.out with per-frame max/min/avg force deviation. Use thresholds:

• max_devi_f < 0.05 eV/Ang: model is reliable
• 0.05 < max_devi_f < 0.15: candidate for active learning
• max_devi_f > 0.15: model is unreliable, do not trust results

Preparing LAMMPS Data File

Convert from CatGo structure to LAMMPS data format:

from ase.io import read, write
# Read pymatgen dict, write LAMMPS data
atoms = read('structure.json')
write('structure.lmp', atoms, format='lammps-data')

Or use dpdata (see data/dpdata/SKILL.md).

Parameter Guidance

Parameter	Typical value	Notes
timestep	0.001 ps (1 fs)	Metal units; can use 2 fs for stiff systems
NVT temp damp	0.1 ps	Nose-Hoover damping; 100x timestep
NPT press damp	1.0 ps	Pressure damping; 1000x timestep
dump frequency	100-1000	Every 100 steps = 0.1 ps
neighbor skin	2.0 Ang	Rebuild neighbor list threshold
run	100K-10M	Depends on property of interest

Common Pitfalls

1. Wrong units — DeePMD pair_style requires units metal (eV, Ang, ps). Never use units real.
2. type_map mismatch — atom types in LAMMPS data file must match the order in the DP model's type_map.
3. Unfrozen model — pair_style deepmd needs a frozen .pb file. Run dp freeze first.
4. Too large timestep — 1 fs is safe; 2 fs may cause energy drift for light elements (H).
5. No equilibration — always equilibrate for 10-50 ps before production run. Discard equilibration data.
6. Memory for large models — GPU memory limits apply. For 1M+ atoms, use CPU or multi-GPU.