Google DeepMind has introduced GenCast, a new AI model that outperforms existing weather forecasting systems. Published in Nature, GenCast represents a significant shift in weather prediction, relying solely on AI to make accurate forecasts.
This follows DeepMind’s earlier AI weather model, NeuralGCM, which incorporated AI and physics-based methods to reduce computing power while achieving similar accuracy to traditional systems. GenCast, however, departs from this by using AI methods alone, functioning similarly to ChatGPT. Instead of predicting the next word in a sentence, GenCast predicts the most likely weather condition.
Trained on 40 years of weather data (1979-2018), GenCast was able to forecast 2019’s weather with greater accuracy than the best current system, Ensemble Forecast (ENS), 97% of the time. It was particularly effective in predicting wind conditions and extreme weather, such as the path of tropical cyclones. Enhanced wind prediction could benefit wind power operations, improving turbine scheduling. Better forecasting of extreme weather can aid in disaster preparedness.
GenCast is not the only AI-driven weather model in development. Nvidia’s FourCastNet, released in 2022, and Huawei’s Pangu-Weather, launched in 2023, also apply AI in weather predictions. Unlike these models, which provide deterministic forecasts with single-value predictions, GenCast uses probabilistic forecasts, offering a range of possible outcomes. For example, it might predict a 40% chance of a low of 30°F or a 60% chance of 0.7 inches of rain.
While GenCast shows promise, it’s not poised to replace traditional meteorology. The model’s reliance on historical weather data limits its ability to predict future conditions in a rapidly changing climate. Additionally, the model must continually be updated with new data, as weather patterns evolve.
DeepMind plans to explore using real-time observational data like wind and humidity for future model testing. Though AI models like GenCast excel at certain tasks, they still struggle with predicting conditions in the upper troposphere and the intensity of tropical cyclones.
Experts suggest that AI will work best in collaboration with human meteorologists. “AI models can provide valuable insights, but human forecasters bring in-depth knowledge and the ability to interpret complex data,” says Aaron Hill, a meteorology professor at the University of Oklahoma. DeepMind’s Price agrees, emphasizing the complementary role of AI and human expertise in improving forecasting accuracy.
