Abstract

Mesoscale ocean dynamics play a critical role in climate systems, governingheat transport, hurricane genesis, and drought patterns. However, simulatingthese processes at high resolution remains computationally prohibitive due totheir nonlinear, multiscale nature and vast spatiotemporal domains. Implicitneural representations (INRs) reduce the computational costs asresolution-independent surrogates but fail in many-query scenarios (inversemodeling) requiring rapid evaluations across diverse parameters. We presentPINROD, a novel framework combining dynamics-aware implicit neuralrepresentations with parameterized neural ordinary differential equations toaddress these limitations. By integrating parametric dependencies into latentdynamics, our method efficiently captures nonlinear oceanic behavior acrossvarying boundary conditions and physical parameters. Experiments on oceanmesoscale activity data show superior accuracy over existing baselines andimproved computational efficiency compared to standard numerical simulations.