Machine Learning Tools for Blast Load Prediction in Obstructed Environments

The assessment of human injuries and structural damage following the detonation of a high explosive requires a comprehensive understanding of the blast load parameters. As explosive events are inherently unpredictable, and key details associated to the charge size, shape, location and material cannot be known a priori, obtaining these parameters often requires probabilistic approaches that feature large batches of numerical models with varying input conditions to embrace this uncertainty.

Machine learning (ML) methods have been shown to rapidly provide accurate predictions for many complex multi-parameter problems in a range of disciplines, including applications featuring blast wave coalescence. However, since ML tools develop their predictive accuracy through a training process that requires data from the problem being modelled, a dependency on potentially costly numerical solvers or physical experiments remains. Furthermore, ML tools are often provided with inputs relating to domain-specific parameters, preventing them from being used beyond the initial problem set, reducing their generality, and thus, requiring the tools to be re-trained when a new scenario is generated.

This thesis introduces two novel methods that independently reduce the impact of costly data collection processes, and prevent the development of tools with limited potential uses. Firstly, when a batch of numerical models are required for probabilistic assessments or training ML tools, any given model can share a number of solution steps with the others. Hence, simulating all domains from birth to termination may result in large amounts of calculation repetition that needlessly increases the overall computation time. The Branching Algorithm (BA) is therefore introduced as a means of mapping data between domains to ensure that calculation steps are only computed once, by identifying when the parameter fields of each model in the batch becomes unique.

Following this, a Direction-encoded Framework for ML tools is developed to enable predictions of blast loading parameters that are based on the surroundings of each point of interest and its position relative to the charge. Through comparisons to a traditional Artificial Neural Network (ANN), provided with global domain inputs, the framework is applied as the Direction-encoded Neural Network (DeNN) to show that the adapted approach enables predictions to be generated in domains with variable sizes and movable obstacles without requiring additional task-specific training.

The computational benefits of BA and the DeNN are then leveraged in a combined analysis, whereby a dataset is incrementally generated using a numerical solver and the BA, and simultaneously used to train the DeNN until a prescribed performance threshold is met. The DeNN then replaces the solver to generate results for any remaining scenarios being evaluated to further reduce the computation time without a detrimental loss of accuracy.

By reducing the time required to conduct a batch analysis, and developing versatile, robust ML based tools, this thesis has shown that complex obstructed environments can be rapidly modelled with consideration of varied geometrical and charge conditions. Results presented throughout the study therefore contribute to the goal of being able to effectively assess the risk posed by a given threat in a probabilistic manner.