Three-dimensional cohesive finite element simulations coupled with machine learning to predict mechanical properties of polymer-bonded explosives

Abstract

Developing multifactorial predictive models for the design of polymer-bonded explosives (PBXs) is of importance for their further application in insensitive munition fields. As a popular method, finite element simulations can provide a reliable prediction, but are laborious and expensive if considering the extensive design parameter space. In light of this challenge, we proposed a coupled strategy that includes machine learning (ML) and three-dimensional cohesive finite element simulation for effciently predicting the mechanical properties of PBXs. The strain rate, particle volume fraction, interface strength, fracture energy, and the binders are considered as the main factors of tailoring the tensile strength of PBXs. To improve the prediction performance, an augmented database of 2500 data sets utilizing GANs neural network were established and then processed to train and test six ML models. The results show the accuracy and generalizability of the low-computational-cost ML models in predicting the mechanical properties of PBX composites. The predicted values from these models are in good agreement with the experimental ones. Feature contribution analysis demonstrates that the tensile modulus and failure strain are most affected by the binders, while the tensile strength are most affected by the fracture energy. Using the above conclusions as design guidelines, we can develop the new PBX formulations according to different mechanical property requirements for their optimal use across insensitive ammunitions. This strategy can be a viable machine-learning-assisted solution to designing PBXs.