NeuralGCM: A new approach to improve weather forecasts | Generative ai Google Tutorial | Google ai Course | Generative ai Healthcare Mckinsey | Turtles AI

Highlights

• NeuralGCM combines physics and AI: It merges atmospheric physics with AI parameterization for more accurate forecasts.
• Effectiveness in short and long term: Competitive for forecasts up to 10 days and promising over decades.
• Limits and advantages: Underestimates extreme events in the tropics but offers realistic climate simulations.
• Future improvements: Integration of individual modules and energy conservation for comprehensive climate modeling

The new frontier of weather forecasting: how NeuralGCM combines physics and AI to improve forecasts

By merging atmospheric physics with the power of AI, NeuralGCM represents a leap forward in climate modeling. Developed by Google’s AI team and the European Centre for Medium-Range Weather Forecasts, this hybrid model promises accurate short and long-term forecasts.

The GCM (General Circulation Model), long considered one of the most reliable tools for weather forecasts and climate studies, is based on physical equations describing known atmospheric processes, supported by empirical parameterizations for less understood phenomena. However, GCMs are now facing increasing competition from machine learning techniques, which use AI to recognize patterns in meteorological data and predict future conditions. Although effective in the short term, these forecasts tend to become inaccurate over longer periods and lack the ability to address the long-term factors needed for studying climate change.

NeuralGCM represents a synthesis of the physical and AI approach. It combines a "dynamical core" that handles large-scale atmospheric convection physics with an AI that parameterizes other meteorological influences. This model is designed to be computationally efficient and has shown excellent performance in weather forecasting benchmarks. Remarkably, NeuralGCM can produce reliable results even over decades, offering the possibility to address some relevant climate issues.

NeuralGCM is a two-part system. The "dynamical core" handles basic physics like gravity and thermodynamics, while everything else, including complex phenomena such as clouds, precipitation, solar radiation, and surface drag, is managed by the AI. This AI is monolithic, trained to handle all these processes simultaneously rather than through separate modules for each phenomenon.

The system is trained jointly rather than separately between the physical core and the AI. Performance evaluations and updates to the neural network are initially performed at six-hour intervals to stabilize the system, then extended to five days. The result is a system competitive with the best available models for forecasts up to 10 days, often outperforming the competition depending on the specific metric used, such as tropical cyclones, atmospheric rivers, and the Intertropical Convergence Zone. Despite the lower resolution compared to purely AI models, NeuralGCM produces less blurry features in long-term forecasts, significantly reducing computational requirements.

However, NeuralGCM has some limitations. It tends to underestimate extreme events in the tropics and does not directly model precipitation, instead calculating the balance between evaporation and precipitation. A specific advantage of NeuralGCM is that it is not limited to short-term forecasts. Tested over two-year seasonal cycles, it successfully reproduced major atmospheric circulation features, including tropical cyclones with realistic trajectories.

One problem encountered is instability in some long-term simulations. These instabilities, often manifested as values going out of bounds, are generally caused by unusual geographical features, such as tall mountains. Despite this, a good percentage of NeuralGCM simulations ran for decades. In a study of 37 simulations started in 1980, 22 lasted for 40 years to the present. The model was also tested with elevated sea surface temperatures, responding well to moderate temperature increases, although the AI struggles with extreme increases expected by the end of the century.

NeuralGCM cannot be considered equivalent to GCM-based climate models. For example, sea temperature changes were manually inserted, and many key factors influencing future climates, such as greenhouse gas concentrations, are not directly reflected in the data used to train the system. However, its ability to predict extreme weather events under current or future conditions makes it a valuable forecasting tool.

Overall, while NeuralGCM has shown significant improvements in weather forecasting compared to traditional models, many aspects still need development. A possible future improvement could consist of replacing the monolithic AI with individual modules focusing on different climate processes. Further integrating the non-AI part of the code to include essential physical processes such as energy conservation may be necessary for a complete climate model.