Overview
DP-GEN (Deep Potential GENerator) coordinates the generation of deep-learning interatomic potential-energy and force-field models. It depends on DeePMD-kit and serves as a workflow layer around molecular simulation and first-principles calculation software, rather than as a standalone pretrained potential. The source describes a concurrent-learning approach that samples structures and selects a subset for first-principles calculations, supporting model development without requiring every sampled structure to be labelled.
The documented entry points cover the main generation workflow, initial-data preparation for bulk, surface and reactive systems, reduction of existing datasets, and Deep Potential testing. Example inputs are provided in JSON format. The intended outputs include initial training data, reduced datasets and generated potential models, depending on the workflow selected. DP-GEN also prepares job scripts, manages HPC job queues and analyzes results, placing it between scientific input preparation and the external programs that perform training or calculations.
The README lists interfaces to simulation packages such as LAMMPS, Gromacs and AMBER, and first-principles packages including VASP, PWSCF and CP2K. It also lists Slurm, PBS, LSF and cloud-machine support. Python 3.10 or newer is required. These are documented interfaces, not evidence that a particular software combination or cluster configuration has been tested here. The source excerpts do not establish system-specific accuracy, convergence criteria or computational cost; those remain matters for evaluation on the intended chemical system.
Key Features
- Concurrent-learning workflow that samples structures and selects a subset for first-principles calculations while developing Deep Potential models.
- Automatic preparation of job scripts, HPC queue management and result analysis, with documented support for Slurm, PBS, LSF and cloud machines.
- Initial-data workflows for bulk, surface and reactive systems through `dpgen init_bulk`, `dpgen init_surf` and `dpgen init_reaction`.
- Interfaces to DeePMD-kit, molecular simulation programs including LAMMPS, Gromacs and AMBER, and multiple first-principles calculation packages.
- Existing-dataset reduction through `dpgen simplify` and Deep Potential testing through `dpgen autotest`.
Use Cases
- Intended evaluation: develop a Deep Potential model for a bulk material using the initial-data and main generation workflows, then assess its suitability for the target simulations.
- Intended evaluation: prepare initial data for a surface or reactive system using the corresponding initialization workflow.
- Intended evaluation: reduce an existing training dataset with `dpgen simplify` and examine whether the reduced dataset remains suitable for the intended model.
- Intended evaluation: assess HPC workflow integration for a documented simulation and first-principles software combination before committing to a larger campaign.
How to Use
- Start with the DP-GEN documentation and choose between model generation, initial-data preparation, dataset reduction and testing. Match the workflow to the scientific task before configuring calculations.
- Prepare a Python 3.10-or-newer environment using the linked environment guide. Follow an installation method in the official README, including its installation check.
- Work through the DP-GEN tutorials and inspect the repository's JSON examples. Consult the initialization documentation if starting without initial data.
- Use the main workflow guide to specify workflow parameters and machine settings for the selected external calculation programs. Evaluate a limited example before scaling up.
- Examine generated data and results. Consult Autotest for potential evaluation, Simplify for dataset reduction, and the user guide for troubleshooting.