deepmodeling/dpgen

DP‑GEN – Automated generation of deep‑learning interatomic potentials

What it is – DP‑GEN (Deep Potential GENerator) is a Python‑based platform that builds accurate, machine‑learning force fields for atomistic simulations. It drives a concurrent learning loop: generate candidate structures, select the most informative ones, run first‑principles (DFT) calculations, and train a Deep Potential model with the DeePMD‑kit.

Why it matters – Traditional force fields require hand‑crafted functional forms and extensive manual tuning. DP‑GEN automates the data‑generation pipeline, can screen millions of configurations while only a few hundred are sent to expensive quantum‑chemistry codes, and produces a uniformly accurate model that can be used in large‑scale molecular dynamics.

Key capabilities

• Scalable workflow – Works on HPC schedulers (Slurm, PBS, LSF) and cloud resources; automatically creates and monitors jobs.
• Broad software integration – Interfaces to MD engines (LAMMPS, GROMACS, AMBER), quantum packages (VASP, Quantum ESPRESSO, CP2K, SIESTA, Gaussian, etc.), and the DeePMD‑kit for model training/inference.
• Modular design – Separate commands for initialization (dpgen init_*), main run (dpgen run), dataset reduction (dpgen simplify), and model testing (dpgen autotest).
• User‑friendly – Installable via pip or conda; a single dpgen -h shows the full CLI; extensive docs, tutorials, and example JSON input files.

Typical workflow

1. Init – Generate an initial configuration set for bulk, surface, or reactive systems (dpgen init_bulk, etc.).
2. Run – Execute the concurrent learning loop (dpgen run), which repeatedly:
   • samples structures with MD,
   • selects uncertain configurations using an ensemble of Deep Potential models,
   • launches DFT calculations on the selected structures,
   • retrains the model with the new data.
3. Simplify – Optionally prune the dataset to a manageable size.
4. Autotest – Validate the final model on independent test sets.

Installation

# via pip
pip install dpgen
# or via conda (recommended for HPC environments)
conda install -c conda-forge dpgen

DP‑GEN requires Python 3.10+ and a working DeePMD‑kit installation.

Getting started – After installing, run dpgen -h to see sub‑commands, then follow the step‑by‑step tutorials at https://tutorials.deepmodeling.com/en/latest/Tutorials/DP-GEN/. Example JSON input files are provided in the examples/ directory.

License & community – Released under the GNU LGPL v3.0; the project is maintained by DeepModeling and welcomes contributions.

---

DP‑GEN is a genuine AI‑driven scientific software package for generating deep‑learning interatomic potentials, fitting squarely within the frontier of machine‑learning‑enabled molecular modeling.