Stay up to date with free updates
Simply sign up to the myFT digest on climate change, delivered straight to your inbox.
Artificial intelligence is helping to make breakthroughs in accurate long-range weather and climate forecasting, according to research that promises advances in both forecasting and the wider use of machine learning.
The team of scientists found that a Google-led model called NeuralGCM, which combines machine learning with existing forecasting tools, successfully combines AI with traditional atmospheric physics models to track multi-decade climate trends and extreme weather events such as cyclones.
The researchers suggest that combining machine learning with established techniques could provide a template for refining the use of AI in other fields, from materials discovery to engineering design. NeuralGCM is much faster than traditional weather and climate forecasting, and outperforms AI-only models at long-term predictions, the researchers say.
“NeuralGCM shows that combining AI with physics-based models can dramatically improve the accuracy and speed of atmospheric climate simulations,” said Stephan Heuer, a senior staff engineer at Google Research and co-author of a paper on the work published in Nature.
According to the paper, NeuralGCM proved to be faster, more accurate, and use less computing power when tested against current forecast models based on an atmospheric physics tool called X-SHiELD, which is being developed by a division of the National Oceanic and Atmospheric Administration.
In one test, NeuralGCM identified nearly as many tropical cyclones as a traditional extreme weather tracker and achieved twice as many as X-SHiELD. In another test based on 2020 temperature and humidity levels, the error rate was 15-50% lower.
According to the paper, NeuralGCM’s calculations were able to generate 70,000 days of simulations in 24 hours using one of Google’s customized AI tensor processing units. In comparison, in a comparable calculation, X-SHiELD generated only 19 days of simulations, requiring 13,824 computer units to run it.
Google developed NeuralGCM in collaboration with the European Centre for Medium-Range Weather Forecasts (ECMWF), an intergovernmental organization.
The European group released the model in June, and Google has made the NeuralGCM code open access, using machine learning powered by 80 years of observational data and reanalyses from ECMWF.
Last year, Google’s DeepMind division unveiled an AI-only weather forecasting model called GraphCast that outperformed traditional methods in forecasting up to 10 days ahead.
Established forecasting organisations such as the UK Met Office are also running projects to integrate machine learning into their operations.
Peter Dueben, head of Earth system modelling at ECMWF and co-author of the latest paper, said AI-only models have “often been met with skepticism” by experts because they are not based on mathematical formulas derived from physics.
Combining physics-based and deep learning models “seems to give us the best of both worlds,” he said, adding that the approach is a “big step towards machine learning climate modelling.”
Dueben said there is still “work to be done,” such as enabling neural GCMs to estimate the impact of increasing CO2 on Earth’s surface temperatures. Other areas in which models need improvement include their ability to simulate unprecedented climate events, according to the paper.
Cedric M. John, head of data science for environment and sustainability at Queen Mary, University of London, an expert not involved in the study, said there is “compelling evidence” that neural GCMs are more accurate than machine learning alone and faster than “full physics” models. While there is still “room for improvement,” he suggested the potential for error is measurable and should be upgradeable.
“The key is that this hybrid model captures the ensemble of forecasts well, and the practical meaning of that is that it allows us to derive estimates of the uncertainty in the forecasts,” John said.
Google is becoming increasingly involved in environmental monitoring efforts: The company provides technical support for satellite missions that track global-warming methane emissions and has partnered with the US space agency NASA to help local governments monitor air quality.
Climate Capital
Where climate change meets business, markets and politics. Read the FT’s coverage here.
Interested in FT’s commitment to environmental sustainability? Find out more about our science-based targets here.
