According to Nature ML methods are starting to outperform NWP without supercomputers https://www.nature.com/articles/d41586-023-03552-y and the coolest thing is that they already released the code https://github.com/google-deepmind/graphcast so I wonder how long would it take before we start seeing people pull data from home weather stations to run an hyperlocal fine-tuned model.
The bitter lesson strikes again.
Background! I’m a meteorologist with a background in model dev (among other things) who took the nowcasting model DeepMind developed (Deep Generative Model for Radar, or DGMR) and am trying to apply it to nowcasting satellite using geostationary satellite observations and incorporating numerical weather prediction (NWP) data. I re-wrote the entire thing in TF/Keras and have about 98% of the original functionality – couldn’t get the Self-Attention layer to work, got fed up, and just defaulted to Keras’s stock MHA layer. It’s been a challenge, but I’m happy with the progress I’m making as essentially a one-man band. I love the fact that DGMR’s arch allows train-small/inference-large domain considerations and the number of input priors is independent of the number of forecast outputs. That said, if anyone can explain to me why using Gaussian noise at the lowest level of the U-net allows for generalization to a larger domain for a cGAN, I’d love to read it.
Now, for GraphCast. The results are impressive, and I think the “climate change” questions are largely overblown if the weather models are truly learning physics. It’s not as if the basic laws of Newtonian physics all of a sudden don’t apply if the earth gets warmer modestly. It is an ERA-5 emulator. That is good.
Now, the local forecast question. With the exception of idealized simulations, all regional models in operations are forced with global model output as boundary conditions. Even the flagship regional model of the US, the High Resolution Rapid Refresh (HRRR), is forced by the GFS. So to do a regional forecast for a NN, you’d have to force the boundaries. Else, you just accept nonsense and need to create a domain large enough such that boundary effects don’t affect your localized regional forecast – this is closer to MetNet-3’s use case. Furthermore, if you want to do nose-bleed resolutions at short time horizons (<6 hours), you’re better off just sticking with the nowcasting ConvRNN-GAN/Diffusion architectures. Lastly, the representation of physics and physical processes at 0.25-degree resolution vice 1-km (for radar) or 2-km (for satellite) grid spacing is very, very different.
On the other hand, if you can force the edges of the regional NNs with a larger external forcing, then you’re at par with the current operational NWP solutions.
As a parenthetical, if you want to see another company doing hi-res global simulations, you can check out Jua.ai.
Is Jua actually doing hi-res global simulations, though? I still have yet to see any actual model outputs, preprints, or peer-reviewed literature from the technical team there.