mat-sample-pes-by-md

What problem does it solve?

This Skill generates diverse, physically meaningful atomic configurations by sampling off-equilibrium potential energy surfaces (PES) with molecular dynamics (MD) and then selecting representative structures via latent-feature clustering.

Core Features & Use Cases

• Off-equilibrium MD sampling: Runs short MD trajectories at a chosen temperature to explore non-equilibrium configurations for MLIP/PES dataset augmentation.
• Crystal-feature extraction for clustering: Extracts crystal/latent features from MatGL or MACE models to represent each saved MD frame.
• Representative structure selection: Uses clustering (e.g., K-Means; plus small-dataset fallbacks) to downselect a smaller, diverse set of configurations for fine-tuning and benchmarking.
• Supercell expansion and stabilization: Automatically expands to a target atom count (with safety caps) and relaxes before MD to reduce instability from residual stress.
• Practical outputs for ML workflows: Writes sampled CIF structures and records metadata (sample indices and MD step positions), enabling downstream DFT labeling or MLIP training.

Quick Start

Use the mat-sample-pes-by-md skill to sample off-equilibrium PES configurations from LiMnO2 by running the provided run_sampling.py script on your input CIF with MatGL or MACE, setting total steps, temperature, number of clusters, and an output directory.