Beyond training data: how elemental features enhance ML-based formation energy predictions

Abstract

Quantum mechanics (QM) based modeling allows for accurate prediction of molecular and atomic interactions, enabling simulations of many materials and chemical properties. However, the high computational cost of QM models leads to a need for faster computational methods to study atomic-scale interactions. Graph Neural Networks fit to QM calculations have been used as a computationally efficient alternative to QM. Still, generalization to diverse unseen compounds is challenging due to the many possible chemistries and structures. In this work, we demonstrate the effectiveness of utilizing element features in facilitating generalization to compounds containing completely new elements in the dataset. Our findings show that we can even randomly exclude up to ten percent of the elements from the dataset without significantly compromising the model's performance.