Using machine learning to predict the ballistic response of structures to projectile impact

Abstract

Ballistic loading is a primary risk in both civil and military defence applications, where successfully predicting the dynamic response of a material to impact is a fundamental component of the design of safe and fit-for-purpose protective structures. Approaches to understand the response to ballistic impact conventionally revolve around experimental tests, whereby the material or structure of interest is subject to impact by a projectile across a range of impact velocities. However, experimental testing is expensive and incurs large costs due to the destructive nature of the testing and the specialist equipment required. Numerical tools, such as the Finite Element (FE) method, play an important role by filling the gaps left sparse by experimental results and contribute towards the complete dynamic material characterisation campaign. This thesis considers an alternative to FE models by using Machine Learning (ML) techniques that learn directly from the available ballistic data. Specifically, the thesis considers the use of Multi-Layer Perceptron (MLP) models to predict the ballistic response of multi-layered targets to impact but its primary intention is to explore the value that generative networks can bring to the ballistic domain. This thesis shows how Generative Adversarial Networks (GANs) can be used to supplement sparse ballistic datasets by generating new samples representative of the dataset that it was trained on, but also how they can be used to predict key ballistic parameters for engineering design such as the ballistic limit velocity, vbl. And finally, how conditional-GANs (cGANS) can be utilised to allow the network to be conditioned on additional auxiliary information such as class labels that refer to a specific property relevant to the ballistic data thus allowing the cGAN to generate new samples specific to the class label given. This allows the trained cGAN to generate data for classes that are not present in the training set and conduct its own material characteristic campaign. The justification for using ML practices for in the ballistic domain lies in the idea that numerical models are adjusted such that the output is consistent with the results from experimental testing. There is therefore an opportunity for research to explore whether ML techniques can capture that same distribution by training on the ballistic data directly.