Deep Learning-Based Prediction of High Explosive Induced Fluid Dynamics

Abstract

Underwater explosions produce complex fluid phenomena relevant to diverse applications including maritime engineering, medical therapeutics, and inertial confinement fusion. These systems exhibit multiphase flows, chemical kinetics, and highly compressible dynamics that challenge traditional computational approaches. Current hydrodynamic solvers, while accurate, are computationally expensive and non-differentiable, limiting their use in design optimization and real-time applications. Here we show that deep neural networks can predict underwater explosion-induced fluid dynamics 4,025 times faster than traditional solvers while maintaining mean absolute percent errors below 0.005\% across all fluid state variables. Our approach maps from explosive material thermodynamic parameters to the temporal evolution of shock fronts and material interfaces, enabling rapid prediction of system behavior for a broad range of ideal explosive materials. Feature importance analysis reveals that exponential decay parameters of the explosive equation of state are the primary drivers of system dynamics, uncovering a previously unknown relationship between thermodynamic compatibility and energy transfer efficiency at material interfaces. Furthermore, we demonstrate an inverse design framework that leverages the differentiability of our neural surrogate to perform parameter discovery, recovering unknown explosive material properties to within 1\% accuracy through gradient-based optimization. This combination of rapid inference, physical insight, and inverse design capabilities provides a route to engineering controlled fluid behavior in underwater explosive systems through material design. We anticipate our approach could enable new applications in defense systems, underwater manufacturing, and medical procedures where precise control of shock waves and bubble dynamics is essential.