Interesting call for more horsepower behind numerical methods. One of the benefits of the emerging machine learning forecast models is dramatically lower need for computation (and therefore energy).
The motivation here is that numerical weather prediction (NWP) models are used for a lot more than just weather forecasting - they're critical research infrastructure and tools. NWP allows researchers to simulate complex atmospheric flows and phenomena which might not be readily or directly observed - or to manipulate observed flows in ways to evaluate dynamical theory and the underlying physics. You can't really do any of this with the AI emulators, which output a tiny, tiny sliver of information relative to the vast array of data you can dump from an NWP system.
Regarding energy usage, you're right up to a point. Today's ML weather models are trained to emulate very large reanalysis datasets (e.g., ECMWF's ERA5) which are produced using the physics-based models. No one has yet to demonstrate an end-to-end AI/ML weather forecast which sidesteps these reanalysis datasets (in fact, the trend so far has been to include _more_ physics-based model datasets for fine-tuning, including archives of historical NWP forecasts or GCM simulations). So there's a massive sunk cost in energy usage to create those training datasets, and then there's the ongoing energy cost of training AI emulators from the ground up. And of course, there's the future energy cost of running more physics-based models to better support the development of AI emulators.
For pure inference/forecasting? Sure. Much faster to run AI models and much lower energy usage. But that's quite literally the tip of the iceberg when it comes to forecast model development.
Sure, the AI stuff is not a complete solution because it relies on the results of numerical methods. I was only suggesting that in the question of making up the marginal difference between the best and fourth-best weather predictions, there might be some application of these other novel methods.
If the question is can the Americans build a gigantic supercomputer and write fortran programs, the answer to that seems obvious. Yes we can do that when the relevant bureaucracy gets pointed in the right direction.
> One of the benefits of the emerging machine learning forecast models is dramatically lower need for computation (and therefore energy).
Maybe but honestly, it doesn't make a ton of sense to me that something that is so power hungry it needs new power plants is going to save energy.
I'm also skeptical that AI can really model a chaotic system better than a.. chaotic system simulation. AI is a statistical pattern matcher. I'm really struggling to understand how that would be better at prediction than a simulation of the underlying phenomena. The whole thing seems like a solution in search of a problem.
> it doesn't make a ton of sense to me that something that is so power hungry it needs new power plants is going to save energy.
The fact that Sam Altman has a 200MW chatbot is not relevant here. For example the GraphCast system from DeepMind runs in under 1 minute on a single TPU device.
