July 22, 2024

expert reaction to machine learning model for accurate weather predictions and climate simulations

A study published in Nature looks at a machine learning model for weather and climate.

Dr Kieran Hunt, NERC Research Fellow, University of Reading, said:

“In the last couple of years, we have seen a dramatic rise in AI weather forecast skill, such that AI models now surpass traditional “physics-based” models when used to predict weather up to a week in advance. However, these AI models suffer from two common problems due to the way they’re trained: firstly they don’t strictly obey the laws of physics (such as keeping the amount of water in the earth system constant), and secondly they tend to produce increasingly “smooth” forecasts the further out they’re used (leading, e.g., to very weak storms). These issues make them useless for climate projections as they quickly drift towards unrealistic scenarios. The NeuralGCM team have overcome this by creating a hybrid model that combines the best features of both the physics-based and AI models, resulting in a climate model that more closely obeys physical laws, while retaining most of the advantage of speed offered by AI.”

Dr Neven Fuckar, Postdoctoral Research Assistant, University of Oxford, said:

“NeuralGCM is the state-of-the-art hybrid atmospheric model based on a dynamical core – differentiable atmospheric model for resolved physics at coarse horizontal resolutions of 2.8o, 1.4o, and 0.7o – combined with neural network models for unresolved small-scale processes in an atmospheric column (e.g., cloud formation, precipitation, radiative transfer, etc.) that are in most comprehensive atmospheric models represented by semi-empirical parametrisation. NeuralGCM shows higher skill than primarily data-driven atmospheric models based on machine-learning (ML) architecture, and comparable or higher skill than the state-of-the-art dynamical models at higher resolution for medium-range (up to 10 days) weather forecasts to decadal climate predictions. However, this novel hybrid model is much faster than typical dynamical counterparts of similar skill which allows about 1,000 to 100,000 times savings in computational time. This is exciting hybrid dynamic-ML approach that has potential to substantially push the boundaries of current weather forecasting and climate prediction capabilities using already available or near-future planned computational and data storage resources in research and operational institutions around the world.”

Professor Hannah Cloke, climate scientist at the University of Reading, said:

“This is a fascinating study that shows the incredible pace of progress that is being made by using AI in weather and climate forecasts. The authors have taken a clever mixed approach, drawing on the strengths of AI to supplement a traditional big simulation of the atmosphere to make predictions that are comparable to the regular way of doing it. The researchers are using AI to do faster calculations of complex physical processes, like the formation of clouds and rain.

“This is an impressive feat, but in a sense, this paper in Nature is already out of date. The European Centre for Medium-Range Weather Forecasts is already running a sophisticated AI-based ensemble model that has been publicly available since June.

“This is interesting in that it demonstrates how out-of-date traditional journal publishing has become. Even Nature, with its focus on the science zeitgeist, can’t keep up with the cutting edge of advances in AI in weather and climate forecasting. This paper was submitted in November last year. Eight months is like a decade for AI developers.

“Before we get too over-excited about the prospect of perfect forecasts, we should show some humility about the things we still don’t know. We are getting pretty good at forecasting the atmosphere, but our predictions of where water is going on the land are still pretty clunky. You can have the best AI or the biggest supercomputer in the world, and you will still get garbage outputs unless you have better weather observations and ways of feeding all that information into the system.

“Most importantly, we won’t genuinely have early warnings for all unless all the boring but vital systems are in place to tell people when they are in danger and to help them get out of harm’s way.”

Dr Martin Rogers, deputy of the British Antarctic Survey (BAS) AI Lab, said:

“Most computer models to predict global climate perform the worst in the polar regions, even though we know these regions contribute the most to global sea-level rise. At the British Antarctic Survey (BAS), we look forward to reviewing the performance of this new hybrid model in these critical regions. This study takes a hybrid approach, combining traditional climate models with machine learning approaches. This hybrid approach produces sharper outputs than using machine learning alone and contain uncertainty estimates which are essential for policymakers. The model has been rigorously tested against WeatherBench, a dataset developed and used by the entire research community, for comparing the performance of different climate models.

“At the BAS AI Lab we are at the forefront of using AI and machine learning techniques to answer a broad spectrum of questions in the polar regions, from weather and sea ice forecasting, polar operations and automation, to detecting and tracking wildlife and icebergs from space.”

Professor Cédric M. John, Head of Data Science for the Environment and Sustainability, Queen Mary University of London (QMUL), said:

“This article introduces a new category of climate models: neither a full-physics model, nor a full-machine learning model, but a hybrid of the two. By combining the best of machine learning (learning from data without explicit physics) and full-physics model (solving ordinary differential physics equations), the authors demonstrate that their new model performs better than traditional machine learning models at predicting atmospheric weather over a range of time scales (from a day to 40 years). They achieve this with a relatively simple architecture. The articles shows some compelling evidence over a range of atmospheric phenomena that do indeed suggest that this approach is more accurate than machine learning, and faster than full-physics model. Therefore, I think that the conclusions are supported by the data. For sure, there is room for improvement. The spatial resolution of the model is coarser than other similar simulators, and the full-physics model still outperform the hybrid model on most metrics. But the very simplicity of the architecture, where fully connected neural network with skip-connects are used, means that more advanced approaches could be attempted in the future. Importantly, this hybrid model does well at capturing an ensemble of predictions, and the practical implication of this is that an estimate of the uncertainty of the prediction can be derived.”

‘Neural general circulation models for weather and climate’ by Stephan Hoyer et al. was published in Nature at 16:00 UK time on Monday 22 July 2024.

DOI: 10.1038/s41586-024-07744-y

Declared interests

Professor Cédric M. John: None.

Prof Hannah Cloke: unpaid Fellow at ECMWF, cosupervises PhD researchers at ECMWF and her spouse works at ECMWF.

Prof Martin Rogers: None. 

Dr Kieran Hunt: None. 

For all other experts, no response to our request for DOIs was received.