Batch active learning for microstructure–property relations in energetic materials

Abstract

Polymer-bonded explosives (PBX) exhibit complex microstructure–property relationships, particularly in their shock-to-detonation transition (SDT) behavior. Traditionally physics-based simulations to explore these relationships are computationally expensive and time-consuming for a number of reasons. We present a material informatics framework that leverages batch active learning to efficiently investigate the intricate microstructure-macroscopic property relationships for PBX, significantly reducing simulation time. Our framework integrates multi-output Gaussian Process Regression (MOGPR) to capture complex relationships between microstructural features (including void volume fraction, shape, and distribution) and reaction response (characterized by shock pressure and run-to-detonation distance). The batch active learning component efficiently traverses the microstructure space by strategically selecting the most informative microstructures for additional simulations, maximizing information gain while minimizing computational costs. By iteratively refining the MOGPR model with the most informative samples, we accelerate the learning process and improve the predictive accuracy of the microstructure–property relationships. Our results demonstrate rapid model convergence and high predictive accuracy, with $r^2$ scores of 0.97 for both pressure and run distance predictions in leave-one-out cross-validation after only eight iterations. This approach efficiently navigates the diverse microstructure space, uncovering key factors governing the SDT behavior in PBX. It also has the potential to significantly improve the design and optimization of PBX materials, enabling the development of tailored explosives with enhanced performance and safety characteristics.