PESMaker Roadmap

MVP 1: structure generation and VASP task preparation

• Parse a simple YAML input file.
• Read user structures from POSCAR, CIF, and extxyz.
• Generate supercells and small perturbations.
• Create VASP SCF calculation folders from a template.
• Produce a local task manifest.

MVP 2: job submission and result collection

• Support local, Slurm, and PBS job backends.
• Track submitted, running, completed, failed, and parsed jobs.
• Parse VASP energy, force, stress, convergence, and failure metadata.
• Resume interrupted workflows from the manifest.

MVP 3: dataset assembly and quality checks

• Export extxyz and GPUMD NEP train.xyz.
• Deduplicate near-identical structures.
• Detect short bonds, abnormal energies, abnormal forces, and failed labels.
• Generate a dataset report with source tags and split metadata.

MVP 4: training workflows

• Train GPUMD/NEP from generated datasets.
• Train MACE from generated datasets.
• Sample structures with LAMMPS-MACE foundation models.
• Summarize train/validation/test errors.
• Package the final potential with a model card.

Later modules

• CP2K labeler.
• MatterSim, Orb, SevenNet, and additional sampling backends.
• Application recipes for diffusion, thermal transport, alloys, defects, surfaces, adsorption, transition states, and reactions.
• Active-learning loop with uncertainty or committee-based candidate selection.

Position relative to autoplex

autoplex is a strong reference for automated MLIP workflows, especially random-structure-search-driven potential-landscape exploration. PESMaker should avoid becoming a simplified clone. Its target identity is:

known material or task
  -> targeted sampling with foundation potentials
  -> DFT labels
  -> application-specific MLIP

This is different from:

chemical formula
  -> random structures
  -> DFT labels
  -> broad PES model