Transfer-learning for rapid predictions of CHNO energetic materials sensitivity

Heterogenous energetic materials (HEs) play a crucial role in applications such as mining, munitions, and propellants, these materials feature intricate microstructures that directly influence their sensitivity to external stimuli. The sensitivity of HEs is determined by both their chemical decomposition at the molecular level and the physical structure at the microscopic scale. These materials, which include compounds like TATB, TNT, HMX, RDX, and PETN, display complex and distinct behaviors under shock-loading conditions, making it challenging to model and predict their responses accurately. Current methods often require extensive data for reliable predictions, which limits the development of generalized models for a broad range of energetic materials.

In this study, we introduce a novel transfer-learning framework designed to overcome these challenges by enabling cross-species prediction of shock responses with minimal training data. Transfer learning leverages knowledge from one material dataset and efficiently adapts it to predict the behavior of other CHNO species, allowing us to generalize insights across different materials. Specifically, we developed a chemical decomposition model that integrates thermodynamic and kinetic data to predict the decomposition pathways, rates, and thermal stability of energetic materials. By interpreting the kinetic data obtained from this model as sensitivity indicators, we constructed a kinetic equation that simplifies multistep decomposition into a single-step process.

Again, when HEs experience external impacts or shock waves, the structural imperfections give rise to localized regions of intense heating, known as hotspots. The formation and development of these hotspots are critical to understanding how energetic materials ignite and react at the macroscopic scale. Here, we conducted reactive void collapse experiments on five key energetic materials with varying pressure conditions and void sizes to study their shock responses. We used the transfer-learning model trained on the HMX dataset to predict the shock response of other materials, demonstrating strong predictive performance with minimal additional data.

Our findings provide valuable insights into the underlying mechanisms of shock-induced sensitivity in energetic materials. The successful application of transfer learning not only enhances predictive accuracy but also offers a powerful tool for the rapid screening and development on new and improved energetic materials. This research lays the foundation for future advancements in the field of energetic materials by providing a predictive framework that is adaptable, data-efficient, and capable of informing safer material designs for a variety of practical applications.