Deep Learning-Based Prediction of High Explosive Induced Fluid Dynamics
Francis G. VanGessel, Mitul Pandya · 2025
Read Deep Learning-Based Prediction of High Explosive Induced Fluid Dynamics by Francis G. VanGessel, Mitul Pandya (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-14710.txt.
Introduction
Francis VanGessel and Mitul Pandya show in a 2025 study that deep neural networks can predict the fluid motion set off by an underwater blast far faster than standard solvers. Blasts under water create complex flows with shock fronts and mixing layers. Solvers that track them step by step are accurate but slow, and that slowness limits design work and real-time use.
Their network takes the traits of the explosive material as input and predicts how shock fronts and material borders move over time. In tests it ran thousands of times faster than the standard solver while keeping errors in the fluid measures extremely small. The method works across a broad range of ideal explosive materials. It opens a path to quick design sweeps for fields from ship building to medicine.
Catalog record
| CREATOR | Francis G. VanGessel, Mitul Pandya |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 73 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-14710
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Deep Learning-Based Prediction of High Explosive Induced Fluid Dynamics Preprint: AIP/123-QED Francis G. VanGessel, Mitul Pandya 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. I Introduction Underwater high-explosive (HE) materials produce a fluid dynamic response characterized by numerous fluid and thermal phenomena. In particular, these HE-water systems exhibit multiphase flow, discontinuous solution fields at the shock and material interface locations, detonation fronts coupled with regions of rapid chemical kinetics, and highly compressible flows which sample large regimes of thermodynamic state space Cole 1948 . These fluid phenomena extend beyond underwater explosions, manifesting in diverse scientific and engineering domains. Pressure-driven bubble dynamics influence propeller performance and erosion in maritime applications Denner 2024 ; Yusvika et al. 2020 . Chemical kinetics play a crucial role in supersonic and hypersonic aerodynamic flows Bertin and Cummings 2006 , while thermodynamically-driven processes govern energy transfer in inertial confinement fusion Gaffney et al. 2018 . Similarly, shock wave propagation enables the operation of rotating detonation engines Mendible et al. 2021 , while understanding shock and bubble interactions is essential for damage mitigation of maritime structures Tong et al. 2022 . The controlled application of cavitation, another related fluid phenomenon, has enabled advances in medical therapeutics Haskell et al. 2023 . Rapid, accurate, and differentiable neural surrogates of an HE-water system could enable near real-time prediction capabilities Brunton and Kutz 2022 , investigation of the HE material property space (e.g. thermophysical properties) Chen, Chiu, and Fuge 2019 and its influence on resulting fluid dynamics, and inverse design for engineered system behaviors. However, these objectives are not achievable through currently available hydrodynamic simulators that are computationally costly and non-differentiable. In this article, we investigate the use of scientific machine learning to overcome traditional computing challenges and address technology needs through the development of differentiable neural surrogates for predicting HE-produced fluid dynamics phenomenon. Such a capability could address similar simulation challenges faced across multiple domains, offering potential advancements in fields ranging from hypersonic aerodynamics to medical cavitation therapeutics that share these complex fluid physics characteristics. Among the numerous fluid structures present in the HE-water system, the shock and material interface locations hold particular engineering relevance. Namely, the shock location corresponds to the point of maximum pressure and thus dictates the shock-induced damage imparted to nearby structures. Furthermore, it is at the multiphase HE-water interface where mechanical work is done by the high-pressure gaseous bubble on it’s surroundings. These bubble dynamics have the potential to damage nearby maritime structures Chen et al. 2018 or even disintegrate kidney and gall stones at smaller scales Brennen 2015 . However, the discontinuous nature of the shock and the multiphase aspect of the material interface makes these fluid structures particularly challenging to model. Specifically, lack of smoothness restricts applicable numerical algorithms while detonation and multiphase phenomena prevent specifying explicit first-principles governing equations for the bubble dynamics Greif and Urban 2019 ; Toro 2013 . Additionally, existing HE materials exhibit a wide variety of detonation and thermodynamic behaviors, thus leading to large variability in the resulting shock and interface dynamics. This work focuses on overcoming these numerous computational challenges and developing rapid and differentiable surrogate predictor of the dynamic fluid states at the shock and material interface for the most prominent class of HE materials. Computation of the fluid response due to HE detonation has traditionally been performed using hydrodynamics simulations (often termed hydrocodes ) which solve the Navier-Stokes equations in the high-pressure regime assuming adiabatic and inviscid flow Cao et al. 2017 ; Kira, Fujita, and Itoh 1999 , under these assumptions the governing equations are referred to as the Euler equations . Traditional hydrodynamic solvers are hampered by significant computational expense and lack of derivative information between simulation inputs and outputs. Lengthy simulation times are a result of the second-order accuracy of the numerical algorithms used in the vast majority of hydrodynamic solvers Toro 2013 . The spatial order of accuracy in turn necessitates highly refined spatiotemporal discretization. Furthermore, hydrocode numerical methods may rely on assumptions of solution field discontinuity, numerical regularization, and non-linear flux terms, all of which render application of adjoint methods problematic Toro 2013 ; LeVeque 2002 ; Fikl et al. 2016 . Physics-based reduced order models (ROMs) partially address the issue of calculation cost, but require application of a number of simplifying thermofluidic assumptions that lower prediction accuracy Cole 1948 ; Kirkwood and Bethe 1942 ; Zhang et al. 2021a ; Zhang et al. 2021b ; Denner 2024 In addition to the computational challenges associated with hydrocode modeling, HE materials exhibit a wide array of thermodynamic properties. These varied thermodynamic properties arise from differences in composition and mass ratios of the HE ingredients and can significantly alter the types of fluid dynamics produced by different HE systems. This range of thermodynamic properties dictate the HE detonation properties that provide the initial conditions to the fluid response. Alleviating computational costs would enable exploration of the HE thermodynamic property space, revealing insights into the complex interplay of shocks, multiphase interfaces, and thermodynamic effects present in these fluid systems. The link between thermodynamic properties, over which one can exercise a certain degree of control, and the fluid dynamic response invites the use of inverse design techniques for engineered system response. However, lack of differentiability of hydrocodes precludes their use within gradient-based inverse design paradigms. In summary, rapid-inference differentiable prediction methods are needed to address computational and design challenges inherent in traditional hydrodynamic computing methods. Scientific machine learning (SML) techniques offer a possible route to predicting the dynamic fluid response via a computationally efficient and differentiable surrogate model. The SML discipline has experienced rapid growth over the past several years and we highlight only a subset of those research efforts with particular relevance to the work presented here. Physics-informed neural networks (PINNs) are widely used partial differential equation (PDE) and ordinary differential