A deep learning-based surrogate model of nonlinear sloshing subjected to random excitations
- 1 October 2026
- journal article
- Published by AIP Publishing in Physics of Fluids
- Vol. 38 (10)
Abstract
Liquid sloshing under external excitation produces strongly nonlinear free-surface motion that is expensive to simulate and difficult to predict over long times with purely data-driven surrogates, which tend to lose stability during recursive rollout. This study presents a deep learning-based surrogate model for nonlinear sloshing subjected to random excitations using a Lagrangian method. The model partitions the dynamics into an explicitly computed response to external forces and an implicitly learned correction for particle interactions, which reduces the learning burden and improves interpretability while enabling fast and stable rollout. In the tested cases, the accumulated error grows approximately linearly at early steps and then stays within a bounded range around 25 mm. The model applies across coupled multi-degree-of-freedom excitation, different tank geometries and dimensions, and varying filling ratios, while preserving the dominant free-surface dynamics. Further analysis traces this stability to the convolution operator, whose kernel is stored on a fixed grid and evaluated by interpolation. This structure constrains the local perturbation response, yielding a narrow distribution of empirical Lipschitz estimates, with only 7% of the states exceeding the neutral value, compared with 47% for an otherwise identical graph-neural-network baseline using a free multilayer-perceptron kernel. It also reduces the disruption caused by topology switching by more than an order of magnitude, from 0.176 to 0.01. The proposed surrogate offers a fast and reliable route to sloshing prediction and provides a useful reference for building particle-level sloshing surrogates in digital-twin applications.Keywords
Funding Information
- National Natural Science Foundation of China (52371329)
- National Natural Science Foundation of China (52001207)
This publication has 51 references indexed in Scilit:
- Physics-informed latent neural operator for real-time predictions of time-dependent parametric PDEsComputer Methods in Applied Mechanics and Engineering, 2025
- Sloshing reduced-order model trained with Smoothed Particle Hydrodynamics simulationsNonlinear Dynamics, 2023
- A CNN-supported Lagrangian ISPH model for free surface flowApplied Ocean Research, 2023
- Neural network-based reduced-order modeling for nonlinear vertical sloshing with experimental validationNonlinear Dynamics, 2023
- A linearized reduced-order model approach for sloshing to be used for aerospace designAerospace Science and Technology, 2021
- Learning slosh dynamics by means of dataComputational Mechanics, 2019
- Nonlinear sloshing in rectangular tanks under forced excitationInternational Journal of Naval Architecture and Ocean Engineering, 2018
- Generalized drag force for particle-based simulationsComputers & Graphics, 2017
- A new volume-of-fluid method with a constructed distance function on general structured gridsJournal of Computational Physics, 2012
- Reduced order modeling of liquid sloshing in 3D tanks using boundary element methodEngineering Analysis with Boundary Elements, 2009