mat-sample-pes-by-md

To generate off-equilibrium configurations for machine learning interatomic potential training, run ASE-based molecular dynamics on a crystal to sample the potential energy surface, then use scikit-learn clustering on extracted latent crystal features to select diverse structures. The workflow outputs CIF files and metadata for downstream DFT labeling.

The best way to sample a potential energy surface for MLIP dataset augmentation is to run short MD trajectories at a chosen temperature, extract crystal features using MatGL or MACE backends, and downselect representative structures via K-Means clustering to ensure physically meaningful diversity.

Yes, you can use MatGL or MACE calculator backends for crystal feature extraction in MD sampling. The workflow uses these frameworks to extract latent crystal features from saved MD frames, which are then passed to scikit-learn clustering for representative structure selection.

Yes, you need to expand your crystal structure to a supercell before running molecular dynamics for MLIP sampling. The workflow automatically expands the input CIF to a target atom count of approximately 40–70 atoms by default and relaxes the structure to reduce instability from residual stress before MD.

To select representative structures from MD trajectories for MLIP fine-tuning, extract latent crystal features from the saved MD frames and apply clustering algorithms like K-Means via scikit-learn. This downsamples the frames into a diverse set of configurations while recording sample indices and metadata.

When sampling potential energy surfaces for DFT labeling, the workflow outputs sampled CIF structures and records run metadata containing sample indices and MD step positions. These practical outputs enable seamless downstream DFT labeling or direct MLIP training.