Google DeepMind released an AI system that is better than current systems at forecasting the weather. GenCast is the new model. Nature today.
Google launched its second AI-based weather model in the last few months. It published in July details on NeuralGCM – a model which combined AI and physics-based techniques like those found in forecasting software. This model was similar to traditional methods, but it used less computer power.
GenCast relies solely on AI. The system works in the same way as ChatGPT but produces weather conditions instead of words. It starts out with weights or random parameters and then compares the prediction to real-world weather data. Over time, GenCast aligns its parameters with actual weather.
Model was created using 40 years worth of data from 1979 to 2018. It then produced a forecast. It was 97% more accurate in its forecasts than the Ensemble Forecast (ENS), the best current forecast. And it did a better job predicting extreme weather, such as the paths of tropical cyclones, and wind conditions. Wind power is more viable with better wind predictions, as they help operators determine when to turn on or off their turbines. Better estimates of extreme weather conditions can also help plan for natural disasters.
Google DeepMind, however, is not the only major tech company that uses AI for weather forecasting. Nvidia launched FourCastNet by 2022. In 2023, Huawei created its Pangu Weather model based on 39 years’ worth of data. The model produces deterministic predictions, which are those that provide a single value rather than ranges. For example, a forecast for tomorrow’s temperature to be 30 degrees F or rainfall of 0.7 inches.
GenCast differs from Pangu-Weather in that it produces probabilistic forecasts–likelihoods for various weather outcomes rather than precise predictions. The forecast could say, “There’s a 40% probability that the temperature will drop to 30 degrees F” or “There’s a 60% likelihood of rain tomorrow.” These types of analyses help officials plan and understand different weather conditions.
This doesn’t necessarily mean that conventional meteorology is dead. Models are trained to predict past weather, but if they’re applied to the future it could lead to incorrect predictions of a climate that is changing.
GenCast still relies on a set of data like ERA5, an hourly estimation of atmospheric variables dating back to 1940. Aaron Hill is a professor at University of Oklahoma’s School of Meteorology, but was not part of this research. He says that the ERA5 model is based on a physics base.
Meteorologists also use equations based on physics to estimate many of the variables that are present in our atmosphere, but which we cannot directly observe. The estimates combined with observational data are fed into models like GenCast. New data is always required. Ilan Price is a researcher at DeepMind, and a creator of GenCast. He says that a model which was developed up until 2018 would perform worse in 2024.
DeepMind will test future models using wind and humidity data to determine if it’s possible to predict the weather based on observations alone.
AI models are not yet able to forecast many aspects of weather, such as the conditions at upper tropopause. While the model is good at predicting the path of tropical cyclones and their intensity, there are not enough data to train the model.
Currently, the hope is that GenCast will be able to work with meteorologists. Price says that there are actual meteorological professionals who look at the forecast and make judgments. They may also use additional data to confirm a forecast if necessary.
Hill is in agreement. He says that the human ability to combine these components is undervalued by AI systems. Human forecasters can look at a lot more data and distill it to produce really accurate forecasts.