Thu Dec 12 09:49:27 UTC 2024: ## AI Weather Forecasting Outperforms Traditional Systems
**Bangalore, India – December 12, 2024** – A groundbreaking new machine-learning model, GenCast, developed by Google DeepMind researchers, is demonstrating superior weather prediction capabilities compared to established systems in certain scenarios. Published today in *Nature*, the research reveals GenCast’s ability to generate more accurate forecasts than the European Centre for Medium-Range Weather Forecasts’ leading model.
GenCast leverages a diffusion model, similar to AI image generators, creating multiple forecasts to account for atmospheric complexities. Unlike traditional models relying on extensive numerical simulations, GenCast achieves this with significantly reduced computational resources and time – producing a 15-day forecast in just 8 minutes using a single TPU. Training the model required five days on 32 TPUs.
The model, trained on reanalysis data from 1979 to 2018, predicts variables like temperature, pressure, and wind speed with improved accuracy by mitigating the “smoothing” effect often observed in other machine learning approaches. Its probabilistic nature provides a superior estimate of future weather by averaging multiple, less-smooth forecasts.
While GenCast shows promise, researchers emphasize the continued importance of traditional numerical weather prediction and reanalysis data. These remain essential for providing initial conditions and refining the machine learning models. Furthermore, the current iteration of the technology isn’t suitable for climate projections due to differences in data availability, time scales, and the influence of factors like carbon emissions.
Despite these limitations, the study suggests that machine learning has a significant role in the future of weather forecasting. However, the researchers highlight that fundamental physics will remain crucial, particularly in bridging the gap between machine learning capabilities and the complexities of climate projections.
First Piper 