PESMaker Project Brief¶
Name¶
PESMaker: Potential Energy Surface Maker.
Working paper title:
PESMaker: a lightweight foundation-potential-assisted workflow for
application-oriented machine-learning potential dataset generation
Core idea¶
PESMaker starts from user-provided atomistic structures and turns them into DFT-labeled datasets and trained machine-learned interatomic potentials.
The intended path is:
known material or task
-> targeted structure generation and sampling
-> DFT SCF labeling
-> dataset quality checks
-> NEP or MACE training
-> final potential
Target users¶
- battery materials researchers;
- solid electrolyte researchers;
- thermal transport researchers;
- alloy and defect researchers;
- surface and catalysis researchers;
- users who already have VASP/GPUMD/LAMMPS scripts but need an automated, reproducible workflow.
Differentiation from autoplex¶
autoplex is a strong reference and should be cited. Its strongest identity is random-structure-search-driven potential-landscape exploration, especially with AIRSS/buildcell and heavy workflow infrastructure.
PESMaker should be different:
- user-provided structures first;
- physical application recipes first;
- foundation potentials for affordable sampling before DFT labeling;
- NEP/GPUMD and MACE as first-class training targets;
- lightweight local execution first, with Slurm/PBS support;
- database services optional rather than mandatory.
First development target¶
MVP 1 should not try to implement all planned science modules. It should support:
- read one or more initial structures;
- generate supercells and optional perturbations;
- create SCF calculation folders from a user template;
- write a manifest of all generated calculation tasks;
- provide a CLI that can validate configs and prepare workflow stages.
After this is stable, add job submission, VASP parsing, extxyz export, and NEP training.