Towards the development of Machine Learning tools for blast load prediction

As explosive events are inherently unpredictable, probabilistic approaches featuring large batches of models with varying input conditions are becoming more prominent in risk assessments and design. Computational fluid dynamics (CFD) allows for the direct solution of blast wave propagation in complex geometries. However, the parameter-rich calculation process can result in prohibitively high computation times, or the need to analyse only a subset of the problem space, making the development of rapid analysis tools essential. This article presents a new analysis approach that leverages the computational benefits from two studies conducted by the author and colleagues to highlight developments made towards this aim. Starting with the Branching Algorithm (BA), informed data mapping reduces the required computation time of a batch of similar explosive scenarios by determining when each model’s parameter field would deviate from the others in the CFD process. Thus requiring fewer models to run from birth to termination. The dataset being generated by the BA is then used to incrementally train the Direction-encoded Neural Network (DeNN), a novel approach for peak parameter predictions in complex domains, in series. Once the DeNN reaches a prescribed performance threshold, it replaces the CFD models for the remainder of the required batch. Together, these approaches allow for robust assessments of varied geometries to be generated with a reduction in computation time of 80%, and average percentage errors of 10.46%, when compared to using CFD models exclusively for a batch of 20 models.