BLAST-NN: Physics-Informed Deep Learning for Blast Wave Propagation

Blast waves exhibit discontinuous rises in pressure, density, and velocity at the shock front. This characteristic leads to numerical solutions of partial differential equations displaying unphysical oscillations at the shock front, presenting a significant challenge for accurate shock front resolution. To address this challenge, a novel physics-informed neural network (NN) algorithm named BLAST-NN is proposed. This algorithm captures the shock front in blast wave propagation by integrating artificial viscosity and the Rankine–Hugoniot relation. Artificial viscosity is employed to avoid the unphysical oscillations at the shock front. The Rankine–Hugoniot relation is applied to ensure the physical consistency of the shock front. The proposed algorithm is validated through three test cases, including the shock tube problem, planar blast wave propagation, and spherical blast wave propagation. The results indicate that BLAST-NN demonstrates agreement with finite difference method solutions for all cases, achieving mean R2 values of 0.97, 0.98, and 0.92, respectively. Additionally, hyperparameter analysis demonstrates that the network architecture, loss function weights, and the shock front detection interval critically influence the algorithm’s prediction accuracy.