The Shear-Bond Descriptor and Universal Machine-Learning Potential Enable Large-Scale Discovery of Negative Thermal Expansion Materials

Abstract Negative thermal expansion (NTE) materials are crucial for mitigating thermal stresses and designing dimensionally stable components in advanced technologies, yet compounds exhibiting this anomalous behavior remain intrinsically rare. The discovery of novel NTE crystals has traditionally relied on serendipitous empirical observation and physical intuition, severely hindered by the absence of universal design principles. Here, by integrating more than a century of accumulated thermal-expansion data with state-of-the-art machine learning, we identify explainable, macroscopic mechanical descriptors for thermal expansion, namely, the shear modulus (G) and the bonding modulus (Ẽ). These descriptors enable the construction of a universal thermal-expansion landscape based on the shear-bond ratio (G/Ẽ), supporting robust classification between positive thermal expansion (PTE) and NTE materials across diverse structural and chemical spaces. Guided by these universal descriptors and powered by high-throughput, finite-temperature simulations using universal machine-learning interatomic potentials, we theoretically predict 3665 NTE crystals exhibiting wide operating temperature windows. Validating this theory-driven paradigm, the experimental synthesis and characterization of 11 previously unreported compounds directly confirm the robust predictive power and transferability of our framework. These results bridge macroscopic mechanics with microscopic lattice dynamics, transforming the discovery of NTE materials into a systematic, scalable, and data-driven science.