JuliaCon 2026

Ocean general circulation models (GCMs) are used for operational forecasting and climate modeling. They simulate a wide range of physical processes, but suffer from biases and uncertainties due to inherent model limitations and inability to model under-resolved processes. These limitations motivate the use of inverse or machine learning methods to systematically constrain models with data from real world observations or high-resolution model runs, thereby reducing both structural and parametric uncertainties. Gradient-based approaches offer a way to “learn” high-dimensional model input spaces, such as physics-based parameters or neural network weights. Combining these methods leads to the notion of neural GCMs that are fully differentiable through automatic differentiation (AD). We have implemented such a capability in the open source ocean GCM Oceananigans.jl, resulting in DJ4Oceananigans (Differentiable Julia for Oceananigans). This required the enhancement of the AD tool Enzyme.jl in conjunction with the compiler tool Reactant.j, which produces a stable multi-level intermediate representation (MLIR) that renders robust, optimized derivative code executable on a wide range of devices including CPUs, TPUs, and GPUs. We demonstrate the computation of accurate gradients across several Oceananigans configurations, for use in sensitivity tests, parameter estimation, and embedded neural network layers. This work represents a significant milestone toward integrating gradient-based inverse methods and machine learning in ocean modeling in Julia, providing open source tools to improve model calibration, reduce persistent model biases, and characterize model uncertainty. It is useful and accessible for other researchers interested in ocean and climate modeling, as well as those interested in an example of differentiability being successfully applied and utilized within a complex scientific model implemented in Julia.