A GPU-accelerated framework for basin-scale tsunami propagation in early warning applications

Effective tsunami early warning systems rely on fast and robust hazard forecasting. While Nonlinear Shallow Water (NLSW) models accurately predict wave dynamics, their high computational cost on CPUs limits their practical use in time-sensitive early warning contexts. Parallel computing on GPUs offers a solution. Celeris Advent (Tavakkol and Lynett, 2017) implements GPU-accelerated numerical simulation but operates on uniform Cartesian grids that can exhibit Earth-curvature-related distortions in trans-oceanic domains.

This study proposes a numerical framework for basin-scale tsunami propagation modelling using a conservative finite volume method (FVM) on a non-uniform orthogonal grid. Conventional plate-carrée grids are straightforward to implement, while their latitude-dependent physical cell dimensions may introduce anisotropic numerical diffusion and compromise temporal stability. Rather than transforming spherical coordinates into generalized curvilinear equations, the proposed approach constructs a non-uniform orthogonal mesh in physical space and explicitly incorporates cell areas and interface lengths into the conservative finite volume formulation. Within this framework, geometric representation and numerical flux evaluation are treated separately to maintain conservation and stability on non-uniform grids. The NLSW equations are discretized using the second-order Kurganov–Petrova central-upwind scheme (Kurganov and Petrova, 2007), with geometric factors integrated through normalization of flux divergence terms by cell areas and interface lengths, while avoiding explicit metric tensor formulations. Initial tsunami generation is computed using the Okada (1985) model based on seismic fault parameters.

The proposed method is designed for GPU parallel computing architectures and is intended for application to large-scale grid computations relevant to basin-scale tsunami forecasting. Standard tsunami benchmark tests are employed to check that the solver can reproduce key nonlinear processes. The framework is further applied to the 2011 Great East Japan Earthquake tsunami as an illustrative example of trans-Pacific propagation modelling.

Beyond numerical performance, the spatio-temporal tsunami fields produced by the framework may be useful for integration into downstream decision-support environments, such as interactive visualization and virtual-reality-based tools for scenario exploration and training.