Asking Graph Neural Networks the Right Questions for Material Discovery.

The ability of machine learning to extract structure-property relationships from density functional theory calculation has enabled the acceleration of material discovery and characterization. Recent publications continue to bolster ever-increasingly accurate models. However, when these models are used in practice, the model's performance is often quite different than the test scenario suggests. This performance discrepancy can often be attributed to the subtleties of how we pose the question that we train the model to answer. In this talk, I will first describe a data augmentation technique we developed to screen for thermodynamic stability [1]. This data augmentation technique yielded a 3-fold reduction in formation energy prediction error compared to previously published models. The model was designed to change the question we ask during training from "what is the formation energy of this structure?" to "what is the formation energy of this structure once relaxed?". I will then describe a new graph neural network architecture and the subtle changes we made during training that profoundly affected the model's performance. The network uses a novel graph embedding and two disconnected graph neural networks to predict the probability of experimental synthesizability of a structure generated by substitution from a prototype structure. Preliminary results show that this method boasts a significant performance boost compared to compositional-based models.<br/><br/>[1] Gibson, J., <i>et al.</i> <i>npj Comput Mater</i> <b>8</b>, 211 (2022).