A critical examination of robustness and generalizability of machine learning prediction of materials properties

Abstract: Recent advances in ML methods have led to substantial improvement in materials property prediction against community benchmarks, but an excellent benchmark score may not imply good generalization of performance. Here we show that ML models trained on the Materials Project 2018 (MP18) dataset can have severely degraded prediction performance on new compounds in the Materials Project 2021 (MP21) dataset. We document performance degradation in graph neural networks and traditional descriptor-based ML models for both quantitative and qualitative predictions. We find the source of the predictive degradation is due to the distribution shift between the MP18 and MP21 versions. This is revealed by the uniform manifold approximation and projection (UMAP) of the feature space. We then show that the performance degradation issue can be foreseen using a few simple tools. Firstly, the UMAP can be used to investigate the connectivity and relative proximity of the training and test data within feature space. Secondly, the disagreement between multiple ML models on the test data can illuminate out-of-distribution samples. We demonstrate that the simple yet efficient UMAP-guided and query-by-committee acquisition strategies can greatly improve prediction accuracy through adding only 1~\% of the test data. We believe this work provides valuable insights for building materials databases and ML models that enable better prediction robustness and generalizability.