A direction-encoded machine learning approach for peak overpressure prediction in urban environments

The use of Machine Learning (ML) models in blast protection engineering has rapidly expanded in recent years, with various publications applying bespoke algorithms to blast wave propagation, fragmentation and structural response problems. The benefits of using this approach for predicting the effects of urban explosions is driven by the need to comprehensively quantify risk through analysing a significant number of unique threats with limited computational expense. However, due to the presence of complex wave coalescence effects, the current state-of-the-art for predicting blast loads in urban environments using ML, the Direction-encoded Neural Network (DeNN), is only able to predict in domains with a limited number of orthogonally placed rectangular obstacles. Therefore, this paper presents a series of developments to the DeNN that allow the tool to predict more complex domains featuring varied obstacle shapes and positions. This is achieved through novel feature engineering that trains the model to understand how the local environment surrounding a point of interest and the strength of the blast wave that is impacting the point influences the magnitude of the prediction. It is shown that peak overpressure can be predicted with an average error of 16.2 kPa for a randomly generated urban environment that emulates a typical city. Future developments will expand the new approach to predict other variables alongside implementing an improved ML architecture.