Beyond Interpolation: Integration of Data and AI-Extracted Knowledge for High-Entropy Alloy Discovery

Abstract

Discovering novel high-entropy alloys (HEAs) with desirable properties is challenged by the vast compositional space and the complexity of phase formation mechanisms. Several inductive screening methods that excel at interpolation have been developed; however, they struggle with extrapolating to novel alloy systems. This study introduces a framework that addresses the extrapolation limitation by systematically integrating knowledge extracted from material datasets with expert knowledge derived from scientific literature using large language models (LLMs). Central to our framework is the elemental substitution principle, which identifies chemically similar elements that can be interchanged while preserving desired properties. To model and combine evidence from these multi-source knowledge, we employ the Dempster--Shafer theory, which provides a mathematical foundation for reasoning under uncertainty. Our framework consistently outperforms conventional phase selection models that rely on single-source knowledge across all experiments, showing notable advantages in predicting phase stability for compositions containing elements absent from training data. Importantly, the framework intends to effectively complement the strengths of the existing methods. Moreover, it provides interpretable reasoning that elucidates element substitutability patterns critical to alloy stability in HEAs formation. These results highlight the framework's potential for knowledge transfer and extrapolation, offering an efficient approach to exploring the vast compositional space of HEAs with enhanced generalizability and interpretability.