Towards Extrapolation in Deep Material Property Regression

Deep learning methods have yielded exceptional performances in material property regression (MPR). However, most existing methods operate under the assumption that the training and test are independent and identically distributed (i.i.d.). This overlooks the importance of extrapolation - predicting material properties beyond the range of training data - which is essential for advanced material discovery, as researchers strive to identify materials with exceptional properties that exceed current capabilities. In this paper, we address this gap by introducing a comprehensive benchmark comprising seven tasks specifically designed to evaluate extrapolation in MPR. We critically evaluate existing methods such as deep imbalanced regression (DIR) and regression data augmentation (DA) methods, and reveal their limitations in extrapolation tasks. To address these issues, we propose an incredibly simple Matching-based EXtrapolation (MEX) framework, which reframes MPR as a material-property matching problem to alleviate the inherent complexity of the direct material-to-label mapping paradigm for better extrapolation. Our experimental results show that MEX outperforms all existing methods and demonstrates exceptional capability in identifying promising materials, underscoring its potential for advancing material discovery. Code is available at https://github.com/panmianzhi/Matching-based-EXtrapolation.