ml-mlip-benchmark

To benchmark an MLIP, you compute Mean Absolute Error and Root Mean Square Error for energy-per-atom and atomic forces against labeled ground-truth data using a provided script. This generates a results JSON with standard error metrics for rapid accuracy validation.

Generating parity plots for materials simulation involves running a plotting script on your benchmarking results JSON. This produces energy and force parity figures to help you quickly spot systematic bias or outliers in your machine learning interatomic potential predictions.

Yes, evaluating MLIP energy and force errors requires numpy for numerical computations and pymatgen for structural data handling. You must execute the benchmarking script within a conda environment that has these dependencies installed to process your labeled dataset.

Yes, you can use this benchmarking approach for pre-training sanity checks across structures in chemistry and drug-discovery research. It evaluates foundation potential baselines by comparing predicted energy and forces against ground-truth labels in your JSON dataset.

Calculating energy MAE and force RMSE requires a JSON dataset containing structural dictionaries paired with ground-truth energy and force fields. This format matches the standard used by related MLIP training data to ensure accurate model evaluation metrics.

MLIP parity plots fail when the plotting script cannot locate the generated results JSON. You must first successfully execute the benchmark script to save your prediction parity data and metrics, then pass that output file to the plotting script.