Electronic configuration vectors provide transferable elemental representations for machine learning across the periodic table
Year: 2026
Type: preprint
Abstract: Machine learning models typically represent elements as discrete identities, limiting inference for elements absent from training. We introduce the electronic configuration vector (ECV), a physically grounded representation derived from the Aufbau principle that continuously encodes orbital occupancies and energetics across elements. Principal component analysis reveals low-dimensional manifolds that organize elements by atomic number, with extrema corresponding to block transitions. Across adsorption energy, formation energy, and metallicity prediction, ECV yields accurate, lightweight, interpretable models while revealing chemical structure. ECV marginally outperforms the Darwin LLM in metallicity classification on the Matbench benchmark (ROC-AUC 0.9614 vs. 0.9598), despite far lower complexity. It also generalizes to unseen elements, achieving 0.16 eV MAE for perovskite formation energies, compared with 0.39 eV for one-hot encoding and 0.18 eV for elemental property vectors. Unlike one-hot encoding, ECV does not impose elemental orthogonality and, unlike property vectors, is definable for every element. These results establish ECV as a coordinate system for transferable machine learning across the periodic table.
Source: ChemRxiv
Institution: Indian Institute of Science Bangalore
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7207787087