A New Generation of Climate Models

M²LInES: Multiscale Machine Learning in Coupled Earth System Modeling

As we embark on the third year of M²LInES, we want to share our progress and what comes next.

M²LInES’ mission is to improve coupled climate models by reimagining physics model development through innovative use of data and AI. We aim to accelerate the pace of climate model development by learning physics from data with scientific machine learning, and ultimately enhance climate model fidelity and reliability for future projections.

🌊 🧊 💨 As we continue to develop and generalize AI-enhanced models of ocean, sea-ice, and atmospheric processes from data, we can now begin to assess their impact on the large-scale climate in a suite of global model configurations.

🌎 Climate models are known to have stubborn biases (model error relative to observations) due to incorrect representations of unresolved physics. We can now demonstrate in GFDL OM4, Global Ocean and Sea Ice Model at 1/4 degree horizontal resolution:

• ✅ A reduction in upper ocean temperature biases in the summer, through the enhancement of a physics-based parameterization with a data-driven vertical diffusivity profile across many forcing regimes (Sane et al., 2023);

Error in summer mixed layer depth between simulation and observations using a physics-based parameterization (left) and the same physics-based parameterizations enhanced with machine learning (right); there is a net reduction in model error, in particular in the Tropics (Sane et al. 2023).

• ✅ A reduction in sea surface temperature biases in the North Atlantic, often associated with poor Gulf stream dynamics and the interaction between turbulent features and the large-scale flow. This increase in skill is done through a new parameterization learned from data and written via a closed-form equation (Zanna & Bolton 2020). A test run of deep convolutional neural networks (Guillaumin & Zanna 2021) implementation via SmartSim has proven we can use GPUs for inference in GFDL MOM6 for global integrations.

Reduction in error of sea surface temperatures compared to observation in global coupled ocean models due to a new data-driven parameterization (left & middle); implementation of a deep convolutional neural network in a global ocean model (right; Courtesy Andrew Shao, HPE).

Stay tuned for publications and code of ocean parameterizations in MOM6 (efforts led by Alistair Adcroft at Princeton/GFDL).

⧮ In addition to developing physics models of specific processes, we are exploring how best to learn the combined error of physics and numerics. New experiments in which we implement data-driven model corrections to improve seasonal forecast of sea-ice extent and concentration in GFDL ocean-sea-ice model (Gregory et al., 2023) and atmospheric regimes in NCAR CAM (Chapman & Berner, 2023) are showing great promise to improve our modeling capabilities and understand the sources of model errors.

Improved sea-ice (left) and atmospheric (right) states through data-driven model corrections.

In the coming months, we will continue to improve the generalization of these hybrid methods in which we combine data, physics, and machine learning and also test a set of new atmospheric parameterizations of the planetary boundary layer (Connolly & Gentine at Columbia University), moist convection (O’Gorman, Yuval, Mooers at MIT), and sea-ice heterogeneity (Holland, Zampieri at NCAR).

Sign up here for our newsletter for monthly updates.

More about M²LInES

Our project comprises ocean, atmosphere, sea-ice scientists, numerical model developers, and machine learning experts. Our collaboration spans 2 continents, 7 academic institutions, and 3 numerical climate modeling centers.

We aim to reduce biases at the air-sea-ice interface in existing global climate models for reliable seasonal to multidecadal timescale projections, focusing on fundamental ocean, atmosphere, and sea-ice processes.

Two leading sources of errors contribute to climate model biases: missing processes and numerics. The missing or inadequate representation of multiscale ocean, sea-ice, and atmosphere processes (e.g., clouds, mixing, turbulence), are not resolved by the current generation of climate models due to computational limitations. Another error source arises from the climate models' numerics, which include spatial and temporal discretizations and numerical dissipation. These errors can accumulate or compensate for each other, making improving climate models intricate and requiring a range of approaches.

To tackle these biases and reduce the potential sources of error, M²LInES’ strategy is to leverage advances in machine learning & ”interrogate” the data to

1. Develop data-informed, interpretable & generalizable subgrid physics models (ocean, ice, atm);
2. Produce error corrections derived from observational products for climate model components.

By improving model physics, this strategy ensures a more faithful representation of feedbacks and sensitivities under different climates.

Our vision

1. ​​Generate new scientific knowledge in climate science from innovative use of data and machine learning: e.g., which physics did we overlook that might be important for scale interaction?
2. Accelerate end-to-end, from development to delivery, for a new generation of climate models; this includes learning and testing parameterizations in global frameworks to tackle significant biases in climate models.
3. Drive a change of direction in the field by building models and tools centered around data-driven methods for the community to advance climate science discovery.
4. Enable a new generation of versatile scientists working at the interface of machine learning, climate science & numerical modeling.

Laure Zanna, on behalf of the M²LInES team

https://m2lines.github.io/