Learning to generate physical ocean states: Towards hybrid climate modeling

TL;DR

The paper addresses the heavy computational burden of OGCM spin-up and the instability of long-horizon DL emulators. It proposes a hybrid pipeline that uses a denoising diffusion probabilistic model (DDPM) trained on DINO-derived ocean states to generate physically consistent initial conditions, with sampling modified by a gradient term $- \kappa(s) \nabla C(x_s)$ in the update $x_{s-1}= \alpha_s^{-\frac{1}{2}}(x_s - \gamma_s \epsilon_\theta(x_s,s)) - \kappa(s) \nabla C(x_s) + \sigma_s z$. The hydrostatic-balance constraint is defined as $C(x)= \sum_k (\mu_k - \frac{1}{N}\sum_{i,j} x_{ijk})^2$, enforcing the vertical structure observed in training. The evaluation combines a priori physical diagnostics and a posteriori 10-year NEMO integrations, showing reduced density-instability and more stable trajectories, while noting a trade-off between state diversity and physical fidelity.

Abstract

Ocean General Circulation Models require extensive computational resources to reach equilibrium states, while deep learning emulators, despite offering fast predictions, lack the physical interpretability and long-term stability necessary for climate scientists to understand climate sensitivity (to greenhouse gas emissions) and mechanisms of abrupt % variability such as tipping points. We propose to take the best from both worlds by leveraging deep generative models to produce physically consistent oceanic states that can serve as initial conditions for climate projections. We assess the viability of this hybrid approach through both physical metrics and numerical experiments, and highlight the benefits of enforcing physical constraints during generation. Although we train here on ocean variables from idealized numerical simulations, we claim that this hybrid approach, combining the computational efficiency of deep learning with the physical accuracy of numerical models, can effectively reduce the computational burden of running climate models to equilibrium, and reduce uncertainties in climate projections by minimizing drifts in baseline simulations.

Figures (4)

• Figure 1: Pipeline of the training and evaluation protocol. From left to right: training of the diffusion model using a database of stable states produced by our oceanic model, generation of initialization states from our diffusion model and temporal integration using numerical simulation, then evaluation of physical consistency on simulated trajectories.
• Figure 2: Comparison between training data and generated states. Left panel: temperature and salinity fields at two different depth levels from the training data (left) and our diffusion model (right), showing the model's ability to capture complex spatial patterns. Right panel: zonally averaged sections of potential density computed from the temperature and salinity fields, comparing the statistical distribution from the training data (top) with generated samples (bottom).
• Figure 3: Effect of physical constraints on generation and temporal evolution. Top : unconstrained generation; bottom : generation with hydrostatic constraint. Left to right: initial density profiles from generated states, density profiles after 10 years of NEMO integration, spatial variance of sea surface temperature and salinity in generated samples. The constraint successfully realistic stratification at the cost of reduced variability in the generated states.
• Figure 4: Illustration of the areas considered for computing the Bottom Water and Deep Water characteristics. Strength of the Antarctic Circumpolar Current (ACC) and North Atlantic subtropical gyre (NASTG) are also depicted, as they are crucial elements of the global ocean dynamics.