(1/N) autoresearch 🤝 weather forecasting - a thread 🧵
Can an automatic research loop improve a real weather dynamical core by making physics-informed changes? TBH we weren’t expecting much, but the early results were surprising enough for us to share:
No Digg Deeper questions have been answered for this story yet.
autoresearch makes ODE/PDE great again without the neural part : )
no backprop, just evolution
(1/N) autoresearch 🤝 weather forecasting - a thread 🧵
Can an automatic research loop improve a real weather dynamical core by making physics-informed changes? TBH we weren’t expecting much, but the early results were surprising enough for us to share:
(2/N) Our process “dynamaxx” directly optimizes the physical terms of a JAX-differentiable dycore checkpointed from NeuralGCM’s dinosaur backbone. We asked our agent to directly target a fixed set of accuracy metrics, RMSE, MAE, bias across temperature, pressure, geopotential, etc. The picture below is the improvement of the new dycore over dinosaur as a function of experiment index
(10/N) Tagging people who might be interested in this: @shoyer, @FerranAlet, @dkochkov1, @singledispatch, @janniyuval, @IlanPrice, @PeterWBattaglia, @tom_r_andersson, @maxxxzdn
(9/N) Resources: 4x NVIDIA L4s, Codex, Claude Code, WeatherBench2 1.5 degree data
(6/N) We believe this is an incredibly promising research direction for weather forecasting. Many existing ML/AI based weather models often underperform NWP (numerical weather prediction) models in extreme weather scenarios, cyclones, tornadoes, etc. As these events can cause outsized costs/damages, they are obviously top of priority
Ref: Nath & Palmer, Nature 651, 583–584 (2026)
(3/N) Now you may be wondering, how overfit is this? It’s gotta be… right?
The thing is there is **no training or backpropagation** whatsoever. We’ve regularized the search process to only introduce physical priors. We only promote promising candidates to be evaluated on the fixed validation period if they look good over a causal “iteration” dataset.
As a sanity check, we performed a classic exercise in measuring generalization. Depicted is the iteration set improvement plotted against the validation set improvement for all promoted changes. The relationship is essentially linear! The improvement we see on candidate changes in the iteration set is the ~same~ as the validation improvement (to first order). This suggests that our regularization procedure can aid generalization
(7/N) ODE/PDE methods naturally incorporate a stronger inductive bias. For example, the actual fluid continuity equations are modeled. As a result, they can outperform, in situations, pure ML methods in extreme weather events and generating forecasts that are not “over-smooth”
(8/N) We’ll be open sourcing the entire process soon. You’ll find it quite simple!
(4/N) Don’t worry ;) we’ve also reserved a special “golden” out of sample period that we have not touched just to be extra careful of the process
(5/N) Some examples of changes to the dinosaur backbone:
- Stability-aware surface wind diagnostics - Exact Coriolis rotation splitting - Potential-temperature and dry-static-energy transport
(11/N) I thank my advisor @zhaoran_wang for his unparalleled guidance and ideation/vision in this project and my collaborator @jostinjiang for seeing this project through and working on this with me from its infancy
@zhaoran_wang Is there a GitHub?