(152d) Data-Driven Chemical Property Models for Energetic Materials Using Transfer Learning

Due to the hazardous nature of energetic materials, it is useful to have accurate estimates of physical properties related to their handling, such as impact sensitivity and vapor pressure. Unfortunately, many safety-related properties depend on multiscale interactions and cannot be directly computed with high accuracy. By themselves, physics-based property prediction models do not extrapolate well and can fail entirely. While machine learning (ML) can overcome these limitations, ML requires large datasets that are not available for energetic properties. Here, we apply two different transfer learning approaches to predict impact sensitivity and vapor pressure. In the first approach, model parameters are learned to map a chemical graph to properties that can be directly computed, and then these parameters are used to predict impact sensitivity. Specifically, we co-train a directed-message passing neural network on a diverse dataset in order to predict impact sensitivity. In the second approach, we embed a physical model into the neural network to enable extrapolation and improve out-of-sample prediction accuracy for energetic vapor pressures. Our models outperform existing models on a diverse test set and are generalizable.