Dataset underlying the publication "Foundation interatomic potentials learn the practical chemistry they were never taught"

Wat erin zit

This dataset supports a computational study investigating whether the internal atomic representations (embeddings) of universal machine-learning interatomic potentials (MLIPs) implicitly encode established physical-organic and organometallic chemistry, despite never explicitly being trained on such properties. Per-atom embeddings were extracted from three foundation MLIPs (MACE-Omol-0, UMA-s-1p2, ORB-mol) and used as fixed inputs to simple supervised probes (ridge regression and a gated attention-pooling network) trained to recover reference chemical descriptors and observables. The dataset consists of the input molecular geometries (XYZ), the reference label tables and per-atom embedding tensors (HDF5), the cross-validated analysis outputs (metrics, per-point predictions, and control/baseline results as CSV), and the Python data analysis code.

Bestanden en diensten

• DataCite REST API (JSON)api.datacite.org
  Openen