Toward Interpretable Machine Learning Models for Materials Discovery

Abstract

Machine learning and AI methods for modelling important materials properties are now an important technology for rational design and optimization of bespoke functional materials. As effective as these methods are in predicting the properties of new materials, current modelling methods still use efficient but rather arcane molecular features (descriptors) to characterize the materials in the data sets. Data driven machine learning models would be considerably more useful if descriptors that have a degree of chemical interpretability could be used, provided they were also effective at recapitulating the properties of materials in training and test sets used to generate and validate the models. Here we show how a particular type pf molecular fragment descriptor, the signature descriptor, can achieve these joint aims of accuracy and interpretability. We model seven different types of materials properties and compare the performance of models generated from signature descriptors with those generated by efficient but arcane Dragon descriptors. We also show that the key descriptors in the model represent chemical functionalities that make chemical sense, and how these fragments can be mapped back onto exemplar materials to guide chemists in making modifications that will improve their properties.