[ad_1]
Scientific Editorial – Google Research has proposed a new free model that combines physics and machine learning to make short- and medium-term weather forecasts and simulate climate over decades.
Called NeuralGCM, it outperforms some existing weather and climate forecasting models and has the potential to significantly save computing power compared to traditional simulators. A description of it was published in the journal Nature.
Reducing uncertainty in long-term forecasts and estimating extreme weather events is key to helping understand climate change mitigation and adaptation, the journal concludes.
The General Circulation Model represents the physical processes in the atmosphere, oceans and land and is the basis for current weather and climate predictions.
These are physics-based simulators that combine numerical solvers for large-scale dynamics with fine-tuned representations of small-scale processes such as cloud formation.
Machine learning models have been proposed as an alternative approach to weather forecasting, with the advantage of lower computational costs, but they generally underperform atmospheric circulation models for long-term predictions.
In this study, Stephan Hoyer’s Google research team proposed NeuralGCM, an atmospheric model that combines machine learning and physics-based methods.
The authors say the new atmospheric model is “competitive” with the accuracy of 1-15 day forecasts from the European Centre for Medium-Range Weather Forecasts, and compare its design to the X-SHiELD model – which is physics-based.
For forecasts up to 10 days ahead, the accuracy of NeuralGCM is comparable to, and sometimes even exceeds, existing machine learning methods.
NeuralGCM achieved 15% to 50% lower errors than X-SHiELD when predicting global temperature and humidity in 2020, generating these predictions in 8 minutes, compared to 20 days for other physics-based models.
When the authors used NeuralGCM to incorporate sea surface temperatures into 40-year climate projections, they found that the model produced results consistent with the global warming trend observed in data from the European Centre for Medium-Range Weather Forecasts.
It also outperforms existing climate models in predicting cyclones and their paths.
“NeuralGCM produces climate simulations that are of the same accuracy as the best machine learning and physics-based methods,” conclude its managers, who add that these results suggest that machine learning is a viable approach to improving climate models.
Google Research has publicly released NeuralGCM on GitHub, allowing researchers to use it.
NeuralGCM is trained using ERA5, a freely available dataset of dense 3D reconstructions of the Earth’s atmosphere over the past 80 years. It was compiled by the European Centre for Medium-Range Weather Forecasts.
According to Google, the center has expressed interest in using NeuralGCM as part of its experimental set of AI models, although it has not yet been finalized.
In November 2023, the company published another machine learning-based weather forecast model, GraphCast, in Science. The head of DeepMind said it was “significantly superior” to traditional systems and could provide early warnings for extreme weather events. Efei
[ad_2]
Source link
