Prediction of Negative Thermal Expansion in 2D-Materials by Multistep Machine Learning Using Structural Descriptors

The discovery of new negative thermal expansion (NTE) metamaterials can be substantially difficult in experiments and computationally expensive in theoretical investigations. In this study, a multistep machine learning (ML) approach was employed to predict NTE maxima (α_max) and thermal expansion coefficient (TEC) in 2D materials. Our predicted target attributes show high correlation with first-principles calculations within quasi-harmonic approximation (QHA). The key idea is to take structural and experimentally tunable features as input variables, which can predict NTE efficiently. Out of the 234 investigated 2D materials, 194 of them can be labeled as NTE materials, as they show NTE within the considered temperature range, T = 0-1000 K. Blind tests were performed using a set of popular materials selected from the 2DMatPedia database to validate the robustness of the model. This work presents a systematic approach toward the rapid screening and prediction of thermal expansion. The identified important features provide guidance for the future design of new 2D NTE materials.