A new system for weather forecasting and predicting future climate uses artificial intelligence (AI) to achieve results comparable to the best existing models, while requiring much less computing power, according to its developers.
In a paper published today in Nature, researchers from Google, MIT, Harvard University, and the European Centre for Medium-Range Weather Forecasts say their model can “drastically reduce computational costs” and “power large-scale physical simulations that are essential to understanding and predicting the Earth system.”
The NeuralGCM model is the latest in a series of research models that use advances in machine learning to make weather and climate predictions faster and cheaper.
What is NeuralGCM?
The NeuralGCM model aims to combine the best features of traditional models with machine learning approaches.
At its core, NeuralGCMs are so-called “general circulation models” that mathematically describe the physical state of the Earth's atmosphere and solve complex equations to predict what will happen in the future.
But NeuralGCM also uses machine learning, a process that looks for patterns and regularities in huge amounts of data for poorly understood physical processes, like cloud formation. The hybrid approach ensures that the output of the machine learning module is consistent with the laws of physics.
The resulting models can be used to forecast weather days and weeks ahead, as well as climate predictions months and years in advance.
The researchers compared NeuralGCM to other models using a standardized set of forecast tests called WeatherBench 2. For three- and five-day forecasts, NeuralGCM performed roughly on par with other machine learning weather models such as Pangu and GraphCast. For longer-term forecasts beyond 10 and 15 days, NeuralGCM performed roughly on par with the best existing conventional models.
NeuralGCMs have also been very successful in predicting less common weather phenomena such as tropical cyclones and atmospheric rivers.
Why Machine Learning?
Machine learning models are based on algorithms that learn patterns in the data they are fed and then use that knowledge to make predictions. Because climate and weather systems are so complex, training machine learning models requires vast amounts of historical observational and satellite data.
The training process is very costly and requires a lot of computer power, but once the model is trained, using it to make predictions is fast and cheap, which is what makes them so attractive for weather forecasting.
They are expensive to train and cheap to use, just like other types of machine learning models: GPT-4, for example, took months to train and reportedly cost over $100 million, but can respond to queries instantly.
Google Research
The weakness of machine learning models is that they often struggle with unknown situations – in this case, extreme or unprecedented weather conditions. To solve this, the models need to be able to generalize, that is, to extrapolate beyond the data they were trained on.
NeuralGCM appears to outperform other machine learning models in this regard because its physics-based core is rooted in reality. As the Earth's climate changes, unprecedented weather conditions will become more common, and it remains to be seen how well machine learning models will respond.
In fact, no one is using machine learning-based weather models for day-to-day forecasting yet, but this is a very active area of research, and we can be sure that the future of forecasting will involve machine learning anyway.
