MLIPs Ontology: Standardizing Metadata for Machine Learning Interatomic Potentials
A new ontology designed for machine learning interatomic potentials (MLIPs) aims to bring together metadata in this field. MLIPs are useful for calculating quantum-mechanical energies and forces based on DFT or wave-function methods, but they do so more affordably. Despite having various algorithms, datasets, and hyperparameters, the essential metadata for proper comparison and reproducibility is scattered. This ontology, created using OWL 2 DL, includes concepts about methods, hyperparameters, training datasets from DFT, and benchmarks. It has three main parts—Method, Training Data, and Benchmark—and connects with existing materials science and machine learning ontologies, like MDO and ML-Schema. It contains 27 formal classes and properties and was shared on arXiv (2607.23219v1).
Key facts
- MLIPs approximate quantum-mechanical energies and forces from DFT or wave-function methods.
- Metadata for MLIP studies is scattered across papers, scripts, and ad-hoc formats.
- The MLIPs ontology is expressed in OWL 2 DL.
- It captures methods, hyperparameters, training datasets with DFT provenance, and benchmarks.
- The ontology has three modules: Method, Training Data, and Benchmark.
- It connects to MDO, CMSO/ASMO, and ML-Schema ontologies.
- It complements the Croissant dataset schema.
- The ontology declares 27 formal classes and properties.
Entities
Institutions
- arXiv