A computer model that integrates machine learning with traditional weather-forecasting technologies has outperformed existing artificial intelligence (AI)-based techniques in forecasting weather scenarios and long-term climate trends.  

The tool is the first machine learning model to provide accurate ensemble weather forecasts, or ones that present a variety of scenarios. It was first described in Nature on 1st July 22. Its advancement paves the way for forecasting that is faster and less energy-intensive than existing tools, and more detailed than approaches based solely on AI. 

The study co-author Stephan Hoyer, Google Research in Mountain View, California says, “Traditional climate models need to be run on supercomputers. This is a model you can run in minutes”. 

Limitations of General Circulation Models (GCMs) 

General circulation models (GCMs), computer programs that simulate processes in Earth's oceans and atmosphere and predict how they can affect the weather and climate, are a foundation of current forecasting systems. GCMs are based on the rules of physics. However, GCMs are computationally demanding, and machine learning advancements are now starting to offer a better alternative. Hoyer said, “We have terabytes or petabytes (one million times larger than a gigabyte) of historical weather data”. “By learning from those patterns, we can build better models," he added. 

A few machine-learning forecasting models are already available in the market, including GraphCast from DeepMind, with its headquarters in London, and Pangu-Weather from the technology company Huawei, which is based in Shenzhen, China. When it comes to deterministic forecasting, which produces a single weather forecast, these models are about as accurate as normal GCMs. However, GCMs are less trustworthy for long-term climate projections or ensemble forecasting. 

Scott Hosking, who researches AI and environmental data at the Alan Turing Institute in London says, “The issue with pure machine-learning approaches is that you’re only ever training it on data it’s already seen”. “The climate is continuously changing, we’re going into the unknown, so our machine-learning models have to extrapolate into that unknown future. By bringing physics into the model, we can ensure that our models are physically constrained and cannot do anything unrealistic,” he added. 

The hybrid model 

Hoyer claimed that he and his team has developed and trained NeuralGCM, a model that combines “aspects from a traditional physics-based atmospheric solver with some AI components”. They produced climate projections and both short- and long-term weather forecasts using the model. The researchers evaluated NeuralGCM's accuracy by comparing its forecasts with real world data and results from various models, such as GCMs and those that only used machine learning. 

NeuralGCM, like other machine-learning models, might use a fraction of the power needed by GCMs to generate precise short-term, deterministic weather forecasts, one to three days ahead of time. However, when generating predictions for longer than seven days, it produced a lot fewer errors than other machine-learning models. In fact, the ensemble model of the European Centre for Medium-Range Weather Forecast (ECMWF-ENS), a GCM that is generally considered to be the gold standard for weather forecasting, produced long-term forecasts that were comparable to those of NeuralGCM. 

Additionally, the team evaluated the model's ability to predict other weather occurrences, including tropical cyclones. When they compared several pure machine-learning models to both NeuralGCM and ECMWF-ENS, they discovered that the latter provided forecasts that were erroneous and inconsistent. NeuralGCM was also put to the test against global storm-resolving models, which are extremely high-resolution climate models. In a shorter amount of time, NeuralGCM could develop tropical cyclone counts and trajectories that are more realistic. 

Hosking noted that being able to predict such events is “so important for improving decision-making abilities and preparedness strategies”. 

Hoyer and his colleagues are keen to further refine and adapt NeuralGCM. “We’ve been working on the atmospheric component of modelling the Earth’s system … It’s perhaps the part that most directly affects day-to-day weather,” Hoyer says. He adds that the team wants to incorporate more aspects of Earth science into future versions, to further improve the model’s accuracy. 

Source: Nature