dpdata — Format Conversion

When to Use

• User needs to convert DFT calculation outputs to DeePMD training format
• User wants to convert between VASP, QE, CP2K, Gaussian, LAMMPS formats
• User needs to merge, filter, or split trajectory data
• User is preparing training data for machine learning potentials

Prerequisites

1. dpdata installed (pip install dpdata, or python -c "import dpdata")

Supported Formats

Format	Read	Write	Key
VASP OUTCAR	Yes	-	vasp/outcar
VASP POSCAR/CONTCAR	Yes	Yes	vasp/poscar
VASP XML	Yes	-	vasp/xml
QE pw.x output	Yes	-	qe/pw/scf
CP2K output	Yes	-	cp2k/output
Gaussian log	Yes	-	gaussian/log
LAMMPS dump	Yes	Yes	lammps/dump
LAMMPS data	Yes	Yes	lammps/lmp
DeePMD raw	Yes	Yes	deepmd/raw
DeePMD npy	Yes	Yes	deepmd/npy
ExtXYZ	Yes	Yes	extxyz

Workflow Steps

1. Convert VASP OUTCAR to DeePMD

catgo_workflow_engine(action="add_task", params={
  "workflow_id": "wf_xxx",
  "task_type": "shell",
  "name": "convert_data",
  "command": "python convert.py",
  "input_files": {
    "convert.py": "import dpdata\nd = dpdata.LabeledSystem('OUTCAR', fmt='vasp/outcar')\nd.to('deepmd/npy', 'training_data')"
  },
  "system_name": "data_prep"
})

Common Conversion Scripts

VASP to DeePMD (with train/valid split)

import dpdata
import numpy as np

# Load all frames from OUTCAR
d = dpdata.LabeledSystem("OUTCAR", fmt="vasp/outcar")
print(f"Loaded {len(d)} frames")

# Random split: 90% train, 10% valid
indices = np.random.permutation(len(d))
n_train = int(0.9 * len(d))

d_train = d.sub_system(indices[:n_train])
d_valid = d.sub_system(indices[n_train:])

d_train.to("deepmd/npy", "data/train")
d_valid.to("deepmd/npy", "data/valid")
print(f"Train: {len(d_train)}, Valid: {len(d_valid)}")

Multiple OUTCARs to single dataset

import dpdata
from pathlib import Path

d = None
for outcar in Path(".").rglob("OUTCAR"):
    sys = dpdata.LabeledSystem(str(outcar), fmt="vasp/outcar")
    d = sys if d is None else d + sys

print(f"Total frames: {len(d)}")
d.to("deepmd/npy", "merged_data")

QE to DeePMD

import dpdata
d = dpdata.LabeledSystem("relax.out", fmt="qe/pw/scf")
d.to("deepmd/npy", "training_data")

LAMMPS dump to ExtXYZ

import dpdata
d = dpdata.System("dump.lammpstrj", fmt="lammps/dump",
                  type_map=["Ti", "O"])
d.to("extxyz", "trajectory.xyz")

Filter by energy/force

import dpdata
import numpy as np

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

# Remove frames with max force > 10 eV/Ang (likely unconverged)
mask = []
for i in range(len(d)):
    max_f = np.max(np.abs(d["forces"][i]))
    mask.append(max_f < 10.0)

d_clean = d.sub_system(np.where(mask)[0])
print(f"Kept {len(d_clean)}/{len(d)} frames")

CLI Usage

# Quick convert
dpdata convert OUTCAR vasp/outcar deepmd/npy training_data

# System info
dpdata info OUTCAR vasp/outcar

Parameter Guidance

Parameter	Notes
fmt	Format string — must match exactly (case-sensitive)
type_map	Required for LAMMPS formats — maps type indices to element symbols
begin / end / step	Frame selection for large trajectories

Common Pitfalls

1. Missing type_map for LAMMPS — LAMMPS dump files have numeric types, not element names. Always provide type_map.
2. Unconverged frames — VASP OUTCARs may contain unconverged ionic steps. Filter by force magnitude before training.
3. Mixed element order — when merging data from different calculations, ensure consistent element ordering.
4. Large memory for big trajectories — dpdata loads all frames into memory. For >10K frames, process in chunks.
5. Units — dpdata converts to eV/Angstrom internally. LAMMPS real units are auto-converted.