Building Chemical Property Models for Energetic Materials from Small Datasets using a Transfer Learning Approach

The destructive nature of energetic materials testing makes their experimental study hazardous, time-consuming, and costly. Due to precautions needed when handling these materials, it is beneficial to have accurate estimates of safety-related properties, such as impact sensitivity, before a material is synthesized as it may help experimentalists avoid synthesis of materials which are destined to not be useful. Unfortunately, impact sensitivity and other safety-related properties, depend in part on macroscale properties and cannot easily be directly computed. While machine learning (ML) can overcome these limitations, ML requires large datasets that are not available for energetic properties. Here, we apply a transfer learning approach whereby model parameters are first learned to map a chemical graph computed to properties before re-training for impact sensitivity. Specifically, we co-train a directed-message passing neural network (D-MPNN) that learns molecule-level features from a large dataset and use these features to predict impact sensitivity. Both characteristics of the computed dataset and model architecture are important to prediction accuracy. Our model outperforms existing models on a diverse test set and is generalizable.