Surrogate modelling strategies for the prediction of near-field blast impulse

The detonation of a high explosive results in the rapid release of energy as the explosive charge undergoes a rapid change in state and is converted into a high pressure, high temperature gas. As the gas expands, the surrounding air is displaced, resulting in a high pressure shock discontinuity -- a shock wave. As this shock wave propagates away from the charge, it can cause severe damage to any structure that it impacts on. Structural blast engineers are tasked with designing infrastructure in a way that it is robust enough to withstand extreme loading, whilst dealing with several constraints such as time, cost and space. Due to the variability in initiation conditions (such as charge shape, charge location, chemical composition of charge and localised point of detonation), and the subsequent variability in loading produced, it becomes impractical to perform numerical simulations or experiments for all possible scenarios, though an understanding of the loading is required to accurately model structural response. Predictive models are therefore required that can predict the blast load parameters of interest (impulse) given certain input parameters that are fast to run and accurate -- predictive models such as these are known as surrogate models.

The blast protection community, when tasked with assessing the viability and safety of structures (the structural response), need an accurate picture of what exactly the loading is. This loading information is composed of both a magnitude and location of a load across a structure, and therefore any predictive approach must predict both these constituent parts of the loading. However, obtaining this loading information, especially within a blast engineering context when the distance between a charge and target is small, is expensive and physical or numerical experiments are costly in both time and money.

Current predictive approaches are severely limited in this regard, in that they do not provide sufficient accurate information, nor are they flexible to handle more than the most simple scenarios. This thesis proposes strategies for surrogate model development in a blast protection engineering context, that allow the rapid evaluation of structural load, given input conditions for a range of scenarios. Furthermore, this thesis demonstrates three applications of strategies that increase the utility of data and knowledge already obtained, that address the fundamental issue of data being expensive to obtain. To achieve this end three approaches are presented: firstly, data transformation procedures, that reduce the dimensionality of the data enabling the use of simpler surrogate models; secondly, the use of directly including known physics into the objective function when model training as a regularisation procedure; and finally, implementing transfer learning by embedding learned knowledge into the architecture of a neural network. These three applications provide statistically significant improvements to model performance and training efficiency, and provide justification to their use in surrogate modelling strategies generally within blast protection engineering.

The results of this thesis should be used to guide surrogate model development for the prediction of peak specific impulse in the near-field for spherical and cylindrical charges. It presents frameworks for creating surrogate models and demonstrates how prior knowledge can be used to improve the performance of surrogate models, or the efficiency when training surrogate models in a new domain, and thereby drastically reducing the need for new data to be obtained. It is shown extensively that machine learning methods can reliably be used in surrogate model development. The findings presented within this thesis have the potential to be implemented into load prediction software which would be of great utility to the blast protection community and insurance industry.