Physical model-guided deep learning approach for explosion load prediction in three-dimensional typical street configurations

Dingkun Luo, Yang Huang, Suwen Chen,
Physical model-guided deep learning approach for explosion load prediction in three-dimensional typical street configurations,
Defence Technology,
2025,
,
ISSN 2214-9147,
https://doi.org/10.1016/j.dt.2025.12.012.
(https://www.sciencedirect.com/science/article/pii/S2214914725004258)
Abstract: Accurate and fast prediction of explosion loads is crucial for structural blast protection design, post-blast damage assessment and emergency response. However, existing methods struggle to balance computational cost with precision for calculating three-dimensional explosion flow fields. This study proposes a physical model-guided deep learning approach for predicting explosion loads in three-dimensional typical street configurations, including L-intersections, T-intersections and crossroads. Low-precision explosion flow fields calculated from the method of images, together with spatial and temporal features of local fields, are introduced as inputs of the convolutional neural network to capture highly nonlinear relationships between data input and pressure field evolution. The influences of key issues, including spatial and temporal features, network architecture and training strategy, on the model’s performance and computational efficiency are discussed in detail. Finally, through comparison with numerical results, the proposed approach has been validated in producing high-precision flow fields and pressure-time histories while requiring less than 10% of the computational time and only 0.2% of the data storage compared to the corresponding numerical solutions. Due to the incorporation of physical model results and local spatiotemporal information, the proposed approach exhibits good interpretability and generalization. This research contributes to the rapid explosion load prediction in urban areas, facilitating swift structural damage assessment and protective design.
Keywords: Street configurations; Method of images; Deep learning; Blast load