AI and climate modeling update

Things are moving very fast.

I wrote two years ago about the role that AI/ML was playing in helping accelerate useful developments in climate models. In a field where literally new possibilities are emerging on a week-by-week basis, that might as well be prehistory.

So it’s worth reviewing what directions have been fruitful, which ones are going slower than expected, which are struggling, and what is going gangbusters that I didn’t even know about back then. This update is in part due to a workshop I attended last week at the Allen Institute in Seattle, but since a lot of what was discussed there is not yet published, I apologize if I’m a little coy about some details.

Emulation for the win (mostly)

The anticipated success of emulation (of whole models or major components) has been continuing – for model calibration/tuning, for scenario filling, attribution, forecasting, and bigger ensembles. This draws in no small measure in the ability to evaluate the results from the ML version directly using the original physics-based model, and so the appropriate degree of confidence that the emulation is doing what it’s supposed to be doing can be assigned. This issue of ‘trust’ comes up a lot in conversations around this topic.

The highest profile emulators are those being used for weather forecasting (e.g. AIFS from ECMWF, and Google WeatherNext 3) which are doing well and sometimes better (on some short-term forecast metrics) than the physics-based models they are emulating. Emulators for climate and longer-time scales (such as Samudra (Dheeshjith et al. (2025)) or ACE or HiRO-ACEv2) are also appearing to do well in limited situations.

There is an effort (AIMIP) that is attempting to benchmark these efforts (at this point only for atmospheric emulators based on ERA5) and interestingly, is finding that the emulators work well in the domain for which they were designed (i.e. short term forecasts), but have very diverse and mostly unreliable responses in out-of-sample tests such as what happens if you increase the ocean temperatures uniformly. This is perhaps unsurprising – the emulators are not learning physics that they then apply, they are learning correlations which they assume persist. Indeed, there are now many examples of ML-models that have good in-sample validation failing to maintain conservation of energy (for instance) when given a forcing of some sort.

This seems a generic enough problem that their might be a generic reason why this is so, and there is some very clear work by Elizabeth Barnes (in prep) that diagnoses this (something to look forward to!). There is a broader discussion about how to optimally train these emulators if we want them to respond well to out-of-sample tests. Some results were presented at the workshop suggested that spectrally-complex near-random forcings (for instance in ocean temperatures), despite not being physical, might be better at getting the atmospheric emulators to distinguish what is really causal from what is merely correlation.

To give an example, weather emulators trained on ERA5 will generally recognise that persistence is a pretty good forecast most of the time (as have humans!). That correlation gets baked into the machine learning. But if something sudden happens (say you’d like to know what happens when the ocean temperatures are higher), the emulator will follow persistence rather than warm up as a physics-based model would. But if the training data included a lot more sudden and uncorrelated changes in SST then that might lead the emulator to pay more attention to a physically consistent response as opposed to the time-step to time-step inertia. This would need the physics-based models to be run in very different ways than they currently are, but that is not insurmountable.

However, to be clear, none of these emulators are real climate models yet: they don’t have a climate sensitivity to CO₂ (or most any other forcing except SST), they can’t be used for paleo-climate out-of-sample tests (like the last glacial maximum), and don’t provide any understanding for why things change. Thus the need for physics-based models to continue to be developed and used has not gone away (phew!).

Hybrid models

Physics-based models that successfully include ML-parameterizations remain a little thin on the ground. Despite the promise of faster execution the combination of legacy code issues (tying fortran-based models to python-based ML components is trickier than you might expect) and reliability concerns (ML components sometimes blow up or give nonsense results more often than standard parameterizations), these kinds of models are still mostly in an experimental phase and are (as far as I can tell) not going as fast as expected (feel free to disagree with examples in the comments!).

The new kid on the block: Agentic AI

This is the crazy thing. Two years ago, I did not know what Agentic AI was, and even six months ago, I was not fully aware of what it could do. But there is a real use case in climate modeling that this is more or less perfect for.

As readers are likely aware, until very recently, climate models were written in Fortran and run on CPU-based high performance computers. Some groups have embarked on multi-year multi-million dollar translation efforts (to C++/Kokkos, or CUDA) but most did not because of the time it would take and the extra expense in a time of shrinking budgets. Some scientists had thought about the potential for using LLMs to do this (for instance, Zhou et al (2024) or Lalhou et al (2026)), but while these efforts were interesting, they were limited. However, three months ago, a preprint from Koldunov et al showed that a full ocean model could be quickly and successfully translated to JAX using agents. This provided automatic differentiation (!) and GPU capabilities without compromising the physics. One month ago, the same group used the latest Claude model to translate MITGCM (another ocean model) from Fortran to JAX over a weekend. Software engineers at NASA a few weeks ago took a configuration of the GISS ESM and converted it in 18 hours. This, quite frankly, is fundamentally game-changing.

To be clear, this is not trivial. The agents need to develop there own tests to verify that their translations are correct, they need to be guided in how design the new code base, and in and of itself translation to JAX (or CUDA or Julia) doesn’t guarantee code speed up on GPUs. But once this is in place (and the performance issues put aside for a moment), a task that would have taken multiple people multiple years to complete can be done in under a day. Say 20 person-years to under 1 person-day – a factor of nearly 10,000 times faster (and cheaper).

The questions climate modeling groups are dealing with thus have very dramatically changed from ‘what is your multi-year strategy to move towards hardware agnostic codes’, to ‘what does it even mean to be a climate model center when anyone anywhere can fork a version of their own and run it anywhere’. The broader questions (that are also hitting software engineers in industry or open source projects) about the shift from human to agent-written code are also profound. What does it mean to be scientist who’s main expertise was coding? Can people re-envisage their role as code managers? Should people still want to edit code themselves at all? How do we maintain momentum and community on group projects? These issues will take some time to work through and I’m not at all confident I know where we’ll end up. However, what is clear is that any climate model code development plans that were written more than three months ago, are basically obsolete.

It is conceivable that groups that are currently working on their CMIP7 simulations (for the most part using their legacy codes on standard HPC machines) may be able to augment these simulations with newly-translated code bases very quickly – say in the next few months. We shall see!

Coda

Speaking personally, I am quite critical of the majority of uses for LLMs – in replacing search, plagiarizing creative industries, summarizing your email (FFS!), replacing human decision-making in crucial fields (medical insurance, benefit payments etc.), providing fatal ‘medical’ advice, and creating massive problems in science for journal editors and program managers as they get spammed with hundreds of low-quality AI-generated submissions with hallucinated references. The environmental damage of un-permitted gas-based electricity generation for massively debt-laden data centers is increasingly bad for both local pollution and global climate. And the danger of AI slop overwhelming our ability to even discover what is true or known is very real. However, there clearly are cases (such as described above) where (under controlled conditions) they can dramatically accelerate work that would have been very difficult or impossible to do before. Some of the recent AI results in mathematical proofs may also fall into this category (but some definitely do not!). How we should square this circle is not obvious (though here are some thoughts on how to use Generative AI ethically). It might also be smart to save some research dollars to pick up some second-hand cheap GPUs for science if (when?) the AI industry boom turns to bust.

In the meantime, it’s a brave new world that has such creatures in’t.

References

1. S. Dheeshjith, A. Subel, A. Adcroft, J. Busecke, C. Fernandez‐Granda, S. Gupta, and L. Zanna, "Samudra: An AI Global Ocean Emulator for Climate", Geophysical Research Letters, vol. 52, 2025. http://dx.doi.org/10.1029/2024GL114318