Known Unknowns—Out-of-Distribution Property Prediction in Materials

Designing high-performance materials often requires identifying materials with property values that fall outside the known distribution, specifically, out-of-support (OOS) values. One strategy for finding materials with desired properties is inverse design through conditional generation of materials. A complementary approach is screening candidate materials through property prediction. However, both approaches typically struggle when the desired property is outside of the training target support.<br/><br/>In this work, we aim to learn a predictor that performs zero-shot extrapolation to higher ranges of property values given chemical compositions. Our method utilizes common descriptor-based representations derived from elemental properties, which encapsulate fundamental chemical information that directly influences material characteristics. We employ a transductive approach to predict various properties of inorganic bulk materials and explore its extrapolation abilities. Rather than predicting property values directly from an input composition x', our method learns to make predictions as a function of a training composition x and the difference between their representations d(x',x). We provide an analysis of our approach, showing it detects and leverages trends in the periodic table to make predictions. Our method produces OOS predictions closer to the ground truth distribution, whereas strong baselines fail to make predictions that exceed the training distribution range.