(285f) Do Machine-Learned Formation Energies Enable Accurate Predictions of Compound Stabilities?

Machine learning (ML) models in chemistry and materials science are often tasked with predicting compound formation energies. These formation energies can then be used as inputs to construct phase diagrams and determine the stability of materials. Recent efforts have shown that ML models trained on density functional theory (DFT) calculated formation energies for solid-state materials achieve accuracies comparable to DFT itself (relative to experiment). In principle, these ML models should therefore be able to predict the stability of materials with comparable accuracy to DFT at a miniscule fraction of the computational expense.

In this work, we designed a set of tests to assess whether seven recently published machine learning formation energy models can accurately predict the stability of compounds. These tests include the reconstruction of ~20,000 phase diagrams available in the Materials Project database in addition to a simulated materials discovery problem, where the models are tasked with re-discovering stable compounds in sparse chemical spaces. Our findings suggest that ML models based only on chemical composition are generally not capable of reliably predicting compound stability and sheds insights into why that is. We do show that stability predictions can be improved by the incorporation of structure in the material representation.

In total, this work emphasizes the importance of testing machine learning models on real-world applications to assess their potential for incorporation in the materials discovery pipeline. We provide a set of publicly available tests to help facilitate this effort for future model development.

Preprint – C. Bartel, A. Trewartha, Q. Wang, A. Dunn, A. Jain, G. Ceder, 2020, arXiv 2001.10591

Code – https://github.com/CJBartel/TestStabilityML