Utilising Foundation Models for Energetic Materials Property Predictions

Abstract. The high costs and risks associated with the discovery of energetic materials necessitate the development of statistical models capable of generating property predictions to assess the viability of potential materials. The application of deep-learning methods in this area is limited by the availability of high-quality experimental data sets. Large data sets contain labelled molecules in the scale of hundreds, whereas traditional deep learning methods often require millions. Foundation models, such as UniMol and MolCLR, are models pre-trained on large data sets which allow them to 'understand' basic chemical information. These models can then be fine-tuned for a specific application using small experimental data sets. Off-the-shelf, pre-trained models are designed for property predictions of general materials but their use for energetic materials remain unexplored. Energetic materials typically have uniquely high crystal densities and oxygen balances in addition to containing distinct functional groups that store large amounts of chemical energy, potentially causing complications in the use of generalist models. The purpose of this work is to create a benchmark across publicly available materials foundation models to assess their performance in predicting key energetic properties.