A differentiable solver for phase-resolved nearshore wave modelling

Recent advances in numerical modeling have improved the fidelity of coastal wave-propagation simulations, but the high computational solution cost of the governing partial differential equations has motivated parallel solvers on multicore CPUs and GPUs. Simultaneously, researchers are exploring machine-learning methods to achieve comparable efficiency gains, and automatic differentiation (AD) is being incorporated to enable optimization and inverse analyses. This paper introduces CelerisAI, a Python-based implementation of the Celeris nearshore wave model that delivers high-performance execution on multiple CPUs or a GPU and interoperates with machine-learning workflows. CelerisAI exposes the forward simulation to AD, yielding a differentiable solver suited to complex tasks such as inverse problems and data assimilation. We evaluate CelerisAI on standard benchmarks and demonstrate AD-enabled data-assimilation applications.