Mutual information informed novelty estimation of materials along chemical and structural axes

Abstract

Assessing the novelty of computationally or experimentally discovered materials against vast databases is crucial for efficient materials exploration, yet robust, objective methods are lacking. This paper introduces a parameter-free approach to quantify material novelty along chemical and structural axes. Our method leverages mutual information (MI), analyzing how it changes with calculated inter-material distances (e.g., using EIMD for chemistry, LoStOP for structure) to derive data-driven weight functions. These functions define meaningful similarity neighborhoods without preset cutoffs, yielding quantitative novelty scores based on local density. We validate the approach using synthetic data and demonstrate its effectiveness across diverse materials datasets, including perovskites with controlled subgroups, a collection with varied structure types, and predicted lithium compounds from the GNOME database compared against materials in the materials project. The MI-informed framework successfully identifies and differentiates chemical and structural novelty, offering an interpretable tool to guide materials discovery and assess new candidates within the context of existing knowledge.