Abstract:The current explosion in machine learning for climate has led to skilled, computationally cheap emulators for the atmosphere. However, the research for ocean emulators remains nascent despite the large potential for accelerating coupled climate simulations and improving ocean forecasts on all timescales. There are several fundamental questions to address that can facilitate the creation of ocean emulators. Here we focus on two questions: 1) the role of the atmosphere in improving the extended skill of the emulator and 2) the representation of variables with distinct timescales (e.g., velocity and temperature) in the design of any emulator. In tackling these questions, we show stable prediction of surface fields for over 8 years, training and testing on data from a high-resolution coupled climate model, using results from four regions of the globe. Our work lays out a set of physically motivated guidelines for building ocean climate emulators.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to construct a machine - learning model (i.e., an ocean - climate simulator) that can accurately predict the ocean climate. Specifically, the researchers focus on two main issues: 1. **The role of the atmosphere in improving the long - term prediction ability of the simulator**: The researchers explored how to improve the performance of the ocean - climate simulator by providing atmospheric boundary conditions (such as wind stress and surface air temperature), especially when predicting variables such as velocity and temperature. 2. **Representation of multi - time - scale variables**: The researchers also explored how to handle variables with different time - scales (such as velocity and temperature) when designing the simulator to ensure that the model can stably predict the changes of these variables. To address these issues, the researchers used data from a high - resolution coupled climate model, trained a U - Net network, and tested it in four different global regions. The results show that the model can stably predict the changes of the surface field for more than 8 years, thus providing physically meaningful guiding principles for constructing more efficient ocean - climate simulators. ### Specific problem analysis #### 1. The role of the atmosphere in improving the long - term prediction ability of the simulator - **Experimental design**: The researchers tested the cases of not providing atmospheric boundary information, providing only wind stress, providing only surface air temperature, and providing both types of information respectively. - **Result analysis**: - When no atmospheric boundary information is provided, the simulator shows a drift phenomenon in all prediction fields. - Providing wind stress can partially restore the missing dynamics of the velocity field, but it will disrupt the seasonal cycle and has a relatively small impact on temperature prediction. - Providing surface air temperature significantly improves the performance of the simulator due to its high decorrelation time, but it still cannot fully capture the full - time - scale of the velocity field. - When both wind stress and surface air temperature are provided, the simulator performs best in all fields, indicating that high - and low - frequency boundary information is crucial for simulating the surface ocean. #### 2. Representation of multi - time - scale variables - **Experimental design**: The researchers explored how to better capture the dynamics of multi - time - scales by adjusting the length of the training window (i.e., the time step \(\Delta t\) and the number of recursions \(N\)). - **Result analysis**: - The baseline simulator (\(\Delta t = 1\) day, \(N = 4\)) can well capture the velocity dynamics, but performs poorly when simulating the slowly - changing temperature field. - Increasing the time step \(\Delta t\) can significantly improve the accuracy of temperature prediction. For example, when \(\Delta t = 2\) days, the 10 - day Anomaly Correlation Coefficient (ACC) increases from 0.73 to 0.84, and further increases to 0.91 when \(\Delta t = 5\) days. - Increasing the number of recursions \(N\) can also improve the accuracy of temperature prediction. For example, when \(N = 8\) and \(\Delta t = 2\) days, the 10 - day ACC reaches 0.82. - Reducing the training window will lead to a performance decline, especially in terms of temperature prediction. ### Conclusion The researchers demonstrated the importance of atmospheric boundary conditions and the representation of multi - time - scale variables in constructing an efficient ocean - climate simulator through the above experiments. They found that providing wind stress and surface air temperature can significantly improve the performance of the simulator, and appropriately increasing the time step or the number of recursions is helpful for better capturing the slowly - changing temperature field. These findings lay the foundation for constructing long - term ocean - climate simulators in the future.

Building Ocean Climate Emulators

Transfer Learning for Emulating Ocean Climate Variability across $CO_2$ forcing

Samudra: An AI Global Ocean Emulator for Climate

ACE2-SOM: Coupling an ML atmospheric emulator to a slab ocean and learning the sensitivity of climate to changed CO$_2$

ACE: A fast, skillful learned global atmospheric model for climate prediction

Emulation of Cloud Microphysics in a Climate Model

Coupled Ocean-Atmosphere Dynamics in a Machine Learning Earth System Model

A comparison of data-driven approaches to build low-dimensional ocean models

Emulating Fast Processes in Climate Models

LUCIE: A Lightweight Uncoupled ClImate Emulator with long-term stability and physical consistency for O(1000)-member ensembles

Towards Learned Emulation of Interannual Water Isotopologue Variations in General Circulation Models

Diffusion-Based Joint Temperature and Precipitation Emulation of Earth System Models

Probabilistic Emulation of a Global Climate Model with Spherical DYffusion

Emulating mean patterns and variability of temperature across and within scenarios in anthropogenic climate change experiments

A Deep Learning Earth System Model for Stable and Efficient Simulation of the Current Climate

Predicting global patterns of long-term climate change from short-term simulations using machine learning

PACER: Physics Informed Uncertainty Aware Climate Emulator

FaIRGP: A Bayesian Energy Balance Model for Surface Temperatures Emulation

Making Ecosystem Modeling Operational–A Novel Distributed Execution Framework to Systematically Explore Ecological Responses to Divergent Climate Trajectories

Opportunistic Emulation of Computationally Expensive Simulations via Deep Learning

DiffESM: Conditional Emulation of Temperature and Precipitation in Earth System Models with 3D Diffusion Models