Blast pressure-time history super-resolution: A hybrid BiLSTM-Transformer approach with wavelet-based preprocessing

Abstract

Modeling blast pressure effects is challenging due to the dynamic, nonlinear nature of the shock wave. Existing data-driven approaches are inadequate in capturing the complex temporal/frequency characteristics of pressure with simplified geometries and limited temporal modeling. This work presents a hybrid deep learning framework that combines BiLSTM networks with Transformer-based attention to improve predictive accuracy. While BiLSTM captures temporal dependencies, the attention mechanism focuses on critical parts of the signal. Wavelet-based preprocessing enables the network to learn pressure-time patterns by extracting frequency-domain features. Two datasets, including free-field and open-air blasts with barriers, were constructed with physics-based simulations. The model successfully predicts pressure values for finer mesh resolutions not present in the training data, providing high-resolution predictions at low cost. The results demonstrate the potential of the model as a scalable, efficient, and accurate alternative to traditional simulations, and for structural safety and real-time blast assessment.