Google Research has developed a breakthrough hybrid general circulation model (GCM) that combines cutting-edge machine learning components with conventional physics-based techniques to improve weather forecast.
This innovative research on Neural General Circulation Models, which was published in Nature, demonstrates how NeuralGCM may improve weather and climate prediction accuracy beyond that of standalone machine-learning models and traditional GCMs.
NeuralGCM, which was created in collaboration with the European Centre for Medium-Range Weather Forecasts (ECMWF), enhances simulation efficiency and accuracy by fusing ML with conventional physics-based modelling.
Breakthrough in Climate Modelling
Google CEO Sundar Pichai has called it a breakthrough in climate modelling. This is because when compared with the existing gold-standard physics-based models, Google claims that this approach offers weather forecasts that are 2–15 days more accurate.
Besides, it is also capable of reproducing temperatures over the last 40 years more accurately than traditional atmospheric models. Unlike traditional models, NeuralGCM combines traditional physics-based modelling with ML for improved simulation accuracy and efficiency.
According to Stephan Hoyer, an AI researcher at Google Research, NeuralGCM is a combination of physics and AI.
To prove their claim, the researchers used a defined set of forecasting tests called WeatherBench 2 to compare NeuralGCM against other models. NeuralGCM performed comparably to other machine-learning weather models like Pangu and GraphCast for three- and five-day forecasts.
Not The Only One
While NeuralGCM can be called a breakthrough in climate modelling, it isn’t the only one. NVIDIA Earth-2 is a full-stack, open platform that combines physical simulations and machine learning models, like FourCastNet and GraphCast, with NVIDIA’s tools for data visualisation.
However, unlike NeuralGCM, Earth 2 focuses on creating a virtual representation of Earth to quickly and accurately simulate and visualise the global atmosphere.
Then, there is the AI2 Climate Emulator developed by the Allen Institute for Artificial Intelligence (AI2). ACE focuses on quickly mimicking complex climate models using deep learning, allowing researchers to run fast simulations and test climate scenarios efficiently.
Not A Big Achievement
“An important advance in atmospheric modelling and long-term weather prediction, but not necessarily a giant leap in climate prediction.” This was how Texas A&M University atmospheric sciences professor R Saravanan described the findings. Saravanan was not involved in the research.
“The NeuralGCM seems like a significant advancement in pure ML-based modelling at first glance,” Saravanan remarked. “In reality, the opposite is true—the paper emphasises the shortcomings of purely ML-based approaches.”
NASA’s Goddard Institute for Space Studies director Gavin Schmidt said that scientists estimate global heating from greenhouse gases as a range due to climate’s inherent chaos, similar to weather predictions like a “40% chance of rain”.
“Physics-based models can better address this uncertainty by simulating the underlying physics, while AI models, lacking this capability, struggle to account for the inherent unpredictability,” Schmidt added.
He also cautioned about the latest findings, saying machine learning isn’t a replacement for physics. He claimed that because “weather models don’t conserve [things like] energy and water”, “you end up with massive drifts”, which cause systems that are trained on meteorological data to gradually diverge from reality.
Furthermore, Schmidt said that merely using meteorological data to train an AI does not ensure that the results it produces will adhere to these physical bounds.