Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
UP: Fachverband Umweltphysik
UP 5: Modelling and Prediction Across Scales
UP 5.4: Talk
Tuesday, March 10, 2026, 15:00–15:15, MER/0002
Differentiable Operators for Ocean Simulation and Inverse Problems — •Pauleo R. Nimtz1,2, Vadim Zinchenko1, Kubilay T. Demir1, Anthony Frion1, and David S. Greenberg1 — 1Helmholtz-Zentrum Hereon, Geesthacht, Germany — 2Universität Potsdam, Potsdam, Germany
Gradient-based optimization is a powerful tool for fitting mechanistic simulations to observation data and is crucial for modern machine learning frameworks. However, in established numerical models of ocean hydrodynamics (e.g. NEMO, ICON-O, SCHISM) and biogeochemistry (e.g. ECOSMO, HAMMOC), the required gradients with respect to model input are typically not available. To address this, we implemented a family of differentiable operators in PyTorch, focusing on the transport and interaction of tracers in fluid dynamical simulations. We provide key operations such as advection, turbulent mixing and biogeochemical processes, with native GPU support and efficient vectorized code. Gradients are computed automatically with memory-efficient custom backpropagation routines for implicit time integration. These operators enable auto-differentiable simulations, with applications including model tuning, data assimilation, physics-informed neural networks, and hybrid physical and data-driven models. We demonstrate their utility by performing data assimilation to identify initial conditions in simulations of inert tracer transport and spatially extended models of light- and nutrient-dependent aquatic ecosystems. These tools provide an essential step towards building and tuning differentiable ocean models, and fitting them to observation data.
Keywords: Ocean modelling; Data assimilation; Auto-differentiation; Differentiable operators; Inverse problems