Fast, accurate climate modeling with NeuralGCM

NeuralGCM is a new machine learning-based model developed to simulate Earth's atmosphere rapidly and accurately. In collaboration with the European Centre for Medium-Range Weather Forecasts, this model integrates traditional physics-based methods with machine learning to enhance simulation efficiency and accuracy. NeuralGCM generates 2-15 day weather forecasts that outperform existing physics-based models and accurately reproduces temperature data over the past 40 years. Unlike traditional models that rely on simplified parameterizations for small-scale processes, NeuralGCM employs a neural network to learn these processes from historical weather data, improving accuracy.

The model has been trained using ECMWF weather data from 1979 to 2019 and is designed for general atmospheric modeling. NeuralGCM's deterministic model matches the performance of current state-of-the-art models for short-term forecasts and excels in producing ensemble forecasts, outperforming traditional models for 5-15 day predictions. It also shows significant improvements in climate-scale predictions, with a third of the error compared to traditional atmosphere-only models.

NeuralGCM is computationally efficient, being over 3,500 times faster than high-resolution models like X-SHiELD, and can run on standard hardware rather than requiring supercomputers. The source code and model weights are available on GitHub, promoting accessibility for researchers. Future developments aim to incorporate additional climate system components, such as oceans and the carbon cycle, to extend the model's predictive capabilities beyond short-term weather forecasts to longer climate projections.