Blast loading prediction in a typical urban environment based on 3D direction-encoded Bayesian neural network

Accurate and rapid prediction of blast loading in the urban environment is crucial for blast-resistant design of buildings and emergency rescue after explosion. Traditional numerical methods are limited by high computational costs and the large amount of computation time required. This paper proposes a 3D direction-encoded Bayesian neural network (3D-DeBNN) for fast blast loading prediction in the whole 3D space of a typical urban environment. Inspiring by the idea of 2D-DeNN, this paper incorporates a 3D direction-encoded feature engineering to extract physics-informed features characterizing the environmental and physical information around the point of interest (POI) and learns the mapping between these features and blast loading. Bayesian approaches are applied to equip the 3D-DeBNN model with the capability of uncertainty quantification for blast loading prediction, generating MC dropout-based 3D-DeBNN and deep ensemble-based 3D-DeBNN. Validated using simulated explosion data from blastFoam, the MC dropout-based 3D-DeBNN achieves blast loading prediction of the whole 3D urban space in 0.6 s (MAPE 〈 8 %, PICP 〉 80 %), while the deep ensemble-based 3D-DeBNN achieves higher accuracy (MAPE 〈 7 %, PICP 〉 95 %) at 3 s. Compared with the MC dropout-based 3D-DeBNN, the deep ensemble-based 3D-DeBNN achieves better prediction performance, but with some compromises in rapidity. This research provides a novel and effective prospect for blast loading prediction in complex urban environment.