A computer model that combines traditional weather forecasting technology with machine learning has outperformed other artificial intelligence (AI)-based tools Predicting weather scenarios and long-term climate trends exceeded.

The tool, released on July 22nd inNaturewas described 1, is the first machine learning model to generate accurate ensemble weather forecasts – ones that represent a range of scenarios. Its development opens the door to predictions that are faster and less energy intensive than existing tools and more detailed than approaches based solely on AI.

"Traditional climate models have to run on supercomputers. This is a model you can run in minutes," says study co-author Stephan Hoyer, who studies deep learning at Google Research in Mountain View, California.

Current forecasting systems typically rely on general circulation models (GCMs), programs that rely on the laws of physics to simulate processes in the Earth's oceans and atmosphere and predict how they might affect weather and climate. However, GCMs require a lot of computing power, and advances in machine learning offer a more efficient alternative. “We have terabytes or petabytes (a million times larger than a gigabyte) of historical weather data,” says Hoyer. “By learning from these patterns, we can build better models.”

There are already some machine learning forecasting models such as Pangu-Weather, created by technology conglomerate Huawei based in Shenzhen, China, and GraphCast by DeepMind with headquarters in London. These models have similar levels of accuracy to typical GCMs for deterministic forecasting – an approach that generates a single weather forecast. However, GCMs are not as reliable for ensemble forecasts or long-term climate predictions.

“The problem with pure machine learning approaches is that you only ever train it on data it has already seen,” says Scott Hosking, who conducts research on AI and environmental data at institutes in London. "The climate is continually changing, we are heading into the unknown, so our machine learning models need to extrapolate into this unknown future. By incorporating physics into the model, we can ensure that our models are physically limited and cannot do anything unrealistic."

Hybrid model

Hoyer and his team developed and trained NeuralGCM, a model that “combines aspects of a traditional physics-based atmospheric solving method with some AI components,” says Hoyer. They used the model to produce short- and long-term weather forecasts and climate projections. To evaluate NeuralGCM's accuracy, researchers compared its predictions to real-world data as well as the output of other models, including GCMs and those based purely on machine learning.

Like current machine learning models, NeuralGCM could produce accurate short-term, deterministic weather forecasts – between one and three days in advance – while using a fraction of the energy required by GCMs. However, it made far fewer errors than other machine learning models when producing long-term forecasts beyond seven days. In fact, NeuralGCM's long-term forecasts were similar to the forecasts of the European Center for Medium-Range Weather Forecast (ECMWF-ENS) ensemble model, a GCM widely considered the gold standard for weather forecasting.

The team also tested how well the model could predict various weather phenomena, such as tropical cyclones. They found that many of the pure machine learning models produced inconsistent and inaccurate predictions compared to both NeuralGCM and ECMWF-ENS. The researchers even compared NeuralGCM to high-resolution climate models known as global storm-resolving models. NeuralGCM was able to produce more realistic tropical cyclone numbers and trajectories in a shorter time.

The ability to predict such events is “so important for improving decision-making skills and preparedness strategies,” says Hosking.

Hoyer and his colleagues want to further refine and adapt NeuralGCM. “We've been working on the atmospheric component of modeling the Earth system... It's perhaps the part that most directly impacts everyday weather,” Hoyer says. He adds that the team would like to incorporate more aspects of earth science in future versions to further improve the accuracy of the model.