A large atomic partition model for materials discovery

Abstract

Accurate and efficient prediction of benchmark properties is essential to the discovery of diverse functional materials, but searching vast element combinatorial and bonding configurational spaces presents formidable challenges to current computational techniques. Here, we devise a large atomic partition (LAP) model featuring a scheme to partition material properties into constituent atomic attributes, which are validated by a data-driven calibration procedure and assigned to elements across the periodic table, then utilized as raw ingredients to assemble and assess targeted properties of new materials. Distinct subtypes are designated for each element based on local atomic environments such as coordination number and valence state, and the parameter count of the LAP model can be tuned widely to tailor prediction accuracy and computational efficiency. As demonstrative case studies, we explore volumetric cohesive energy, bulk modulus, and shear modulus, and the results showcase superior accuracy, efficiency, universality, and interpretability of the LAP model compared to alternative approaches. Moreover, based on the predicted elastic moduli, we discover a series of rare and highly sought-after compounds exhibiting concurrent superior hardness and toughness, highlighting the promise of the LAP model in high-throughput screening for advanced materials with targeted outstanding functionalities.