Known Unknowns: Out-of-Distribution Property Prediction in Materials and Molecules

Abstract

Developing high-performance materials and molecules often requires identifying those with property values that fall outside the known distribution. Therefore, the ability to extrapolate to out-of-support (OOS) property values is critical for both solid-state materials and molecular design. Given the chemical compositions of solids or SMILES of molecules and their property values, our objective is to learn a predictor that extrapolates zero-shot to higher ranges. In this work, we employ a transductive approach to property prediction and achieve more accurate predictions, as well as a 3x and 2.5x improvement in True Positive Rate (TPR) of OOS materials and molecules identification, respectively. We leverage analogical input-target relations in the training and test sets, enabling generalization beyond the training target support.