While by no means perfect, AI tools for predicting hurricanes and other major storms are both developing quickly and improving rapidly, offering weather forecasters new methods they can consult and rapidly re-run. With offerings from Nvidia, Huawei, and Google being deployed, forecasters are already finding these tools capable of matching conventional atmospheric models under many conditions.
The upside of these AI-based is enormous, however, with frequently much faster rates of execution. Says Mark DeMaria, atmospheric scientist at Colorado State University and a retired division leader for the US National Hurricane Center, of Nvidia’s FourCasetNet model, “If you have a model that’s already trained, our FourCastNet model runs in 40 seconds on a junky old graphics card. So you could do like a whole gigantic ensemble that would not be feasible with physically based models.” The end result is the ability to regenerate projections rapidly as new data comes in, or to plot larger sets of possible scenarios and determine most likely paths.
There are still risks with these models though, as the models are unable to estimate levels of uncertainty in themselves – in many cases, these AI-models are essentially black boxes, where it’s not always clear how the model is arriving at its final projections, whereas this level of uncertainly can be better estimated for conventional models. So, while AI weather forecasts aren’t quite ready to fully take over, they do provide an interesting path for the future.
Source:
https://arstechnica.com/science/2023/09/ai-is-getting-better-at-hurricane-forecasting/amp/
