Machine Learning Based Mesh Movement for Non-Hydrostatic Tsunami Simulation

Yezhang Li, Stephan C. Kramer, Matthew D. PiggottarXiv 2026 · DOI · arXiv

TL;DR

This paper introduces a deep learning-based mesh movement strategy (UM2N) to accelerate non-hydrostatic tsunami simulations. By replacing costly traditional mesh adaptation solvers with a neural network surrogate, the researchers achieved a staggering 290x speedup in mesh adaptation while maintaining the accuracy and robustness needed to model complex coastal wave dynamics.

What Problem Does This Paper Address?

Simulating a tsunami as it approaches the coast is a computational "balancing act." To be accurate, models must account for non-hydrostatic effects—the vertical accelerations that occur as waves steepen and interact with the seafloor. While these effects are vital for predicting run-up and inundation, they are computationally expensive.

Standard fixed-mesh models struggle here: a mesh fine enough to capture the wave peak across an entire ocean is too slow, while a coarse mesh misses the critical physics. Mesh adaptivity (moving the "grid" to follow the wave) is the solution, but traditional methods like solving the Monge–Ampère (MA) equation are themselves slow and prone to "mesh tangling"—where the grid becomes so distorted the simulation crashes. This paper seeks to replace these fragile, slow solvers with a robust, lightning-fast machine learning surrogate.

Methods

The authors integrate a Universal Mesh Movement Network (UM2N) into the Thetis coastal ocean modeling framework.

Key technical components include:

  • The Physics: A depth-integrated non-hydrostatic model implemented using Discontinuous Galerkin Finite Element (DG-FE) methods.
  • The ML Architecture: UM2N uses a Graph Transformer encoder to extract features from the current mesh and a Graph Attention Network (GAT) decoder to predict how nodes should move.
  • r-Adaptivity: Unlike "h-adaptivity" (which adds/removes nodes), this method uses "r-adaptivity," which moves existing nodes to areas of high interest (like the wave front) while keeping the total number of nodes and their connectivity constant.
  • Monitor Functions: The network is guided by a "monitor function" based on the Hessian (curvature) of the water surface elevation. Essentially, the mesh "huddles" where the wave is changing most rapidly.

Datasets

  • Training: The UM2N was trained on a PDE-independent dataset. Instead of using real tsunami data, it was trained on thousands of randomly generated Gaussian fields. This allows the model to learn the general "math" of mesh movement without being tied to a specific physical scenario.
  • Evaluation Benchmarks:
    • N-wave strip source: An idealized tsunami propagation test.
  • Truncated conical shoal: Testing wave refraction over underwater topography.
  • Conical island: A laboratory benchmark (Briggs et al., 1995) involving complex run-up and a wet-dry interface.
  • Monai Valley: A 1/400 scaled laboratory experiment of a real-world Japanese valley, used to test inundation and "run-up" on complex terrain.

Key Insights and Ideas

  • Decoupling Physics from Adaptation: Because the Monge–Ampère equation is an auxiliary "helper" equation, the neural network doesn't need to know the underlying fluid dynamics. It just needs to know how to distribute points based on a provided "importance map" (the monitor function).
  • Superior Robustness: In the "Conical Island" test case, the traditional MA solver failed (diverged) due to mesh tangling within 100 steps. The UM2N model, however, remained stable throughout the entire simulation, proving it is more "physically valid" in extreme conditions.
  • Local vs. Global Refinement: UM2N tends to be more "aggressive," concentrating nodes very locally around steep gradients, whereas traditional MA methods distribute nodes more smoothly. This localized focus is particularly effective for tracking sharp tsunami fronts.

Key Results

  • Massive Speedup: On an Nvidia A100 GPU, the UM2N inference was ~290x faster than the traditional CPU-based MA solver. Even when running on a standard CPU, the ML approach was ~6x faster.
  • High Accuracy: In the Monai Valley benchmark, UM2N achieved a 91.14% error reduction in simulating the wave peak compared to a fixed coarse mesh, matching the accuracy of a much more expensive fine-mesh simulation.
  • Efficiency: The approach allows researchers to achieve "fine-mesh" results using a "coarse-mesh" budget, making probabilistic hazard assessments (which require thousands of simulation runs) much more feasible.

Limitations and Future Work

  • 2D vs. 3D: The current work is limited to 2D depth-integrated models. The authors suggest extending the framework to 3D mesh geometries in future iterations.
  • Boundary Constraints: Currently, the model keeps boundary nodes fixed. Allowing these nodes to move along the boundary could further improve mesh quality and reduce distortion near the edges of the domain.
  • Unsupervised Learning: The authors point toward exploring unsupervised mesh movement networks (like UGM2N) to further enhance the generalizability of the tool across different types of coastal engineering problems.

Summary updated

Abstract

This study investigates the use of machine learning based mesh adaptivity, specifically mesh movement methods (UM2N), with depth integrated non-hydrostatic shallow water models. Motivation for this comes from the need for models which balance efficiency and accuracy for use in probabilistic coastal hazard assessment. Implementations are built on the discontinuous Galerkin finite-element (DG-FE) based software, Thetis, which leverages the partial differential equation (PDE) framework Firedrake for automated code generation. Verification on benchmark test cases and validation against laboratory measurements of coastal hazards, focusing on tsunami propagation, run-up, and inundation is performed. In these tests, the UM2N-driven meshes help resolve key non-hydrostatic dynamics and yield numerical solutions in close agreement with reference computations and measured data. Numerical results indicate that the UM2N surrogate based approach significantly accelerates conventional mesh movement techniques and has high robustness over long integration periods and under strongly nonlinear wave conditions.
BibTeX
@misc{li2026machine,
  author = {Yezhang Li and Stephan C. Kramer and Matthew D. Piggott},
  title = {{Machine Learning Based Mesh Movement for Non-Hydrostatic Tsunami Simulation}},
  year = {2026},
  eprint = {2603.06152},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.06152},
}
Figure