By: Ruihao Rick Shan

San Jose, California- A new weather forecasting system utilizing AI has achieved results comparable to or better than existing models while also needing less computing power.

In a paper published by a team from Google working with researchers from the European Center for Medium-Range Weather Forecasts says that their AI model offers ” orders of magnitude computational savings” and can “enhance the large-scale physical simulations that are essential for understanding and predicting the Earth system.”

The model is called Neuro GCM and combines traditional models with machine learning. NeuralGCM is a type of “general circulation model,” which uses mathematical descriptions of Earth’s atmospheric conditions to solve complex equations and predict future changes.

Additionally, NeuralGCM incorporates machine learning to identify patterns and regularities in large datasets, particularly for less well-understood physical processes like cloud formation. This hybrid approach ensures that the outputs from the machine learning modules align with the laws of physics, allowing the model to make forecasts days or weeks in advance.”

The researchers evaluated NeuralGCM against other models using a standardized set of forecasting tests known as WeatherBench 2. In terms of three- and five-day forecasts, NeuralGCM performed comparably to other machine-learning weather models like Pangu and GraphCast. For longer-range forecasts, spanning over ten and fifteen days, NeuralGCM achieved accuracy similar to the best existing traditional models.

Machine learning models often struggle in unfamiliar situations, such as extreme or unprecedented weather conditions. For effective performance, a model needs the ability to generalize and extrapolate beyond its training data.

NeuralGCM seems to surpass other machine learning models in this aspect due to its physics-based core, which provides a grounding in reality. As Earth’s climate changes, unprecedented weather conditions will become more frequent, posing a challenge for machine learning models to keep up.

Although machine learning-based weather models are not yet used for day-to-day forecasting, the field is highly active in research. Regardless of the outcome, it is certain that future weather forecasts will incorporate machine learning.