Structure-Agnostic Prediction of the Electronic Density of States with a Chemical Language Model

Abstract

The electronic density of states (DOS) is conventionally computed from a relaxed crystal structure, which is unavailable for compounds that have been neither synthesized nor cataloged. Here we introduce DOSSIER ($\textbf{D}$ensity $\textbf{o}$f $\textbf{S}$tates from $\textbf{S}$to$\textbf{i}$chiometry with $\textbf{E}$ncoder $\textbf{R}$epresentations), a chemical language model that maps elemental composition directly to this spectrum. The encoder is pretrained by cross-modal knowledge distillation from a universal machine-learning interatomic potential; the transfer lowers the error by 11% when only 1,000 training examples are available. On the Mat2Spec benchmark, DOSSIER reaches a mean absolute error of 3.76 states eV$^{-1}$ against 3.64 for the best structure-aware model; on an extended Materials Project dataset, the predicted spectra yield band gaps and $\textit{d}$-band descriptors with useful accuracy. Screening 11,977 binary and 13,251 five-component high-entropy alloy compositions for a $\textit{d}$-projected DOS resembling that of NiPt$_{3}$ places known oxygen reduction electrocatalysts near the top of the ranking.