Predicting Drop‐Weight Impact Sensitivity From Molecular Graphs Using Physics‐Informed Artificial Intelligence

ABSTRACT Predicting the drop‐weight impact sensitivity () of energetic materials accurately and efficiently is a crucial step for the design of safe explosives. Traditional experimental determination of values is costly, time‐consuming, and inherently uncertain. This work presents an approach for predicting of pure molecular explosives directly from two‐dimensional molecular graphs, encoded as simplified molecular‐input line‐entry system (SMILES) strings, using physics‐informed artificial intelligence (AI) models. To address limited experimental drop‐weight impact sensitivity data, we augment our dataset of experimentally measured sensitivities of 625 unique high explosive molecules with physics‐informed synthetic predictions for 1 additional real molecules containing energetic functional groups. Using the existing, publicly available Chemprop software—implementing a message‐passing neural network encoder and feed‐forward neural network predictor—we trained a robust predictive model that improves upon existing models. The proposed approach shows a reduction in predictive error of about 10% over an existing physics‐informed model, which predicts from a hand‐crafted set of descriptors. Our results demonstrate the feasibility and effectiveness of leveraging publicly available chemical databases and physics‐informed models for accelerating the screening of energetic materials based solely on molecular graphs.