When will climatologists give up and just go all-in on pure DL. Their current modelling approach is a dead end and correctly drives a lot of climate(ology) skep

When will climatologists give up and just go all-in on pure DL. Their current modelling approach is a dead end and correctly drives a lot of climate(ology) skepticism. The current approach is to try and simulate everything from physical first principles. They started in the 70s with just the atmosphere, later added the oceans. The problem they have is that despite many claims to the contrary, their models have historically not worked and not generated accurate predictions. There are numerous examples of this like the famous "running hot" problem [0], or that since the millenium west Antarctica has cooled rather than warmed, and that no models predicted this [1]. Lack of detail around clouds are sometimes blamed. Climatologists responded by adding ever more detail to the models, in the hope that with bigger programs the predictions will come good and the (enormous) CIs will shrink. That unfortunately didn't happen. Instead as models became more complex they diverged from each other. We are told the science is settled and only the edges need to be tied up, but in reality the science is getting less settled over time as the simulations get supposedly more accurate. It's not possible for all of them to be right but the literature (and IPCC) often just average all the models together, as if they are all equally reliable. Now we are told that figuring out clouds (still very far from done) isn't enough, and marine microbes are also critical. Where does this end? We can't simulate the full complexity of Earth. And if more detail = less confidence continues to hold, adding this will just make the models even less useful. On top of all that, the models are full of numerical stability bugs. NASA Model E is nearly half a million lines of FORTRAN. If you tell it to simulate a stable pre-industrial climate many runs end up catastrophically diverging to a snowball or a climate hotter than Venus [2]. Very often these bugs are papered over with hacks rather than root caused and fixed. For example, it has a comment saying "this routine makes sure that the temperature remains within reasonable bounds during the initialization process (sometimes the computed temperature iterated out in left field someplace, way outside any reasonable range)". Then they do the same for wind speeds, just using min(max(...)) to constrain that part of the simulation when it enters a feedback loop and would otherwise crater the sim. There are dozens of other serious errors that are patched up at the end with hacks. And then the model has hundreds of tunable free parameters that make it easy to overfit, some of which are explicitly computed backwards like humidity thresholds to ensure that at least some runs remain stable, instead of being set to their actual physical values. The field needs to reset its approach. The AI field has dropped attempting to model intelligence from first principles because it didn't work, and embraced DL methods that can model any function given enough data. Google has been training and using neural nets to do short range weather forecasts. Neural climate models may not yield more accurate results (who says there's enough data) but at least there's a base of relatively solid tooling and methodologies for training, detecting overfitting, validation etc and the implementations are written by pro-level developers. [0] https://www.nature.com/articles/d41586-022-01192-2?error=cookies_not_supported&code=a9c480e4-65fb-4603-a909-612a652d631e https://www.nature.com/articles/d41586-022-01192-2?error=coo... [1] https://journals.ametsoc.org/view/journals/bams/104/6/BAMS-D-22-0153.1.xml https://journals.ametsoc.org/view/journals/bams/104/6/BAMS-D... [2] https://www.netzerowatch.com/content/uploads/2023/07/Eschenbach-Climate-Models.pdf https://www.netzerowatch.com/content/uploads/2023/07/Eschenb...