Predicting drop-weight impact sensitivity directly from molecular structure using physics-informed machine learning

A key constraint for the design of energetic materials is sensitivity, or the resistance to an unintended reaction. The most common measure of sensitivity is the height at which a specified weight dropped on a sample of the energetic material induces a reaction with 50% probability (H50). Drop-weight tests to estimate H50 are limited and noisy. To perform fully automated molecular discovery, predictions must use only the information contained in the molecular structure.



This work illustrates a machine learning method that predicts drop-weight impact sensitivity at state-of-the-art accuracy using only molecular structure, as encoded in simplified molecular input line entry system -- SMILES -- strings. The key to the success of this predictor is the augmentation of limited drop-weight test data (for around 500 unique energetic molecules) with over 10,000 additional molecules. A physics-informed predictor of impact sensitivity is leveraged to generate estimated H50 values for these additional molecules. Utilizing the open source Chemprop package, a message-passing neural network is trained to encode molecules into a continuous latent space. Then, a feed-forward neural network is employed to predict H50 values from these continuous encodings. The resulting H50 predictions are at least as accurate as current benchmarks in the literature.