Element selection for functional materials discovery by integrated machine learning of elemental contributions to properties

Abstract

The unique nature of constituent chemical elements gives rise to fundamental differences in materials. Assessing materials based on their phase fields, defined as sets of constituent elements, before specific differences emerge due to composition and structure can reduce combinatorial complexity and accelerate screening, exploiting the distinction from composition-level approaches. Discrimination and evaluation of novelty of materials classes align with the experimental challenge of identifying new areas of chemistry. To address this, we present PhaseSelect, an end-to-end machine learning model that combines representation, classification, regression and novelty ranking of phase fields. PhaseSelect leverages elemental characteristics derived from computational and experimental materials data and employs attention mechanisms to reflect the individual element contributions when evaluating functional performance of phase fields. We demonstrate this approach for high-temperature superconductivity, high-temperature magnetism, and targeted bandgap energy applications, showcasing its versatility and potential for accelerating materials exploration.