Exploring negative thermal expansion materials with bulk framework structures and their relevant scaling relationships through multi-step machine learning.

Discovering new negative thermal expansion (NTE) materials is a great challenge in experiment. Meanwhile, the machine learning (ML) method can be another approach to explore NTE materials using the existing material databases. Herein, we adopt the multi-step ML method with efficient data augmentation and cross-validation to identify around 1000 materials, including oxides, fluorides, and cyanides, with bulk framework structures as new potential NTE candidate materials from ICSD and other databases. Their corresponding coefficients of negative thermal expansion (CNTE) and temperature ranges are also well predicted. Among them, about 57 materials are predicted to have an NTE probability of 100%. Some predicted NTE materials were tested by the first-principles calculations with quasi-harmonic approximation (QHA), which indicates that the ML results are in good agreement with the first principles calculation results. Based on the comprehensive analysis of the existing and predicted NTE materials, we established three universal relationships of CNTE with an average electronegativity, porosity, and temperature range. From these, we also identified some important critical values characterizing the NTE property, which can serve as an important criterion for designing new NTE materials.