General circulation models (GCMs) form the backbone of weather and climate prediction, utilizing numerical solvers for large-scale dynamics and parameterization of small-scale processes such as cloud formation. Despite continuous improvements, GCMs face significant challenges, including persistent errors, biases, and uncertainties in long-term climate projections and extreme weather events. Recent machine learning (ML) models have shown remarkable success in short-term weather forecasts; however, their usefulness is limited by their lack of stability for long-term predictions and their inability to provide calibrated uncertainty estimates.
GoogleAI proposes NeuralGCM to address the limitations of weather and climate prediction using General Circulation Models (GCMs). Traditional GCMs, which rely on physics-based simulations, are computationally intensive and struggle with long-term stability and accurate ensemble predictions. These GCMs combine numerical solvers for large-scale atmospheric dynamics with empirical parameterizations of small-scale processes such as cloud formation. Machine learning models trained on historical data, such as ECMWF's ERA5, have demonstrated good short-term weather prediction capabilities at low computational cost, but have failed in long-term forecasts and ensemble accuracy.
GoogleAI's NeuralGCM is a hybrid model that combines a differentiable solver for atmospheric dynamics with a machine learning component that parameterizes physical processes. The model aims to leverage the strengths of both traditional GCMs and machine learning approaches to provide stable and accurate predictions over a range of time scales with good computational efficiency.
NeuralGCM integrates a differentiable dynamical core and a learned physics module to predict the impact of unresolved atmospheric processes using neural networks. The end-to-end training approach involves backpropagation through multiple simulation steps, gradually increasing the length of the rollout from 6 hours to 5 days. This method allows the model to account for the interactions between learned physics and large-scale dynamics, improving stability and accuracy.
Experiments were conducted to evaluate the performance of NeuralGCM against best-in-class models such as ECMWF-HRES and the Ensemble Forecasting System, as well as machine learning models such as GraphCast and Pangu. For 1-15 day weather forecasts, NeuralGCM achieves comparable accuracy, with lower errors in probabilistic versions and better ensemble average forecasts. For climate simulations, NeuralGCM accurately tracks climate indicators over decades and simulates emerging phenomena such as tropical cyclones, at significantly reduced computational costs.
In conclusion, NeuralGCM successfully addresses the limitations of both traditional GCMs and pure machine learning models, providing a stable and accurate hybrid approach for weather and climate prediction. By combining a differentiable solver with machine learning parameterization, NeuralGCM powers the large-scale physical simulations essential for understanding and predicting Earth's systems, while providing significant computational efficiency.
Please check paper and detailAll credit for this research goes to the researchers of this project. Also, don't forget to follow us. twitter And our Telegram Channel and LinkedIn GroupsUp. If you like our work, you will love our Newsletter..
Please join us 47,000+ ML subreddits
Check out our upcoming AI webinars here
Pragati Jhunjhunwala is a Consulting Intern at MarktechPost. She is currently pursuing her B.Tech from Indian Institute of Technology (IIT) Kharagpur. She is a technology enthusiast with a keen interest in the range of applications of software and data science. She is constantly reading about developments in various areas of AI and ML.
