the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Potential for Equation Discovery with AI in the Climate Sciences
Abstract. Climate change and Artificial Intelligence (AI) are both attracting great interest across society. There is also substantial interest in merging the two sciences, with evidence already that AI can identify earlier precursors to extreme weather events. There are a range of AI algorithms, and selection of the most appropriate one maximizes the amount of additional understanding extractable for any dataset. However, most AI algorithms are statistically based and even with careful splitting between data for training and testing, they arguably remain as emulators. Emulators may make unreliable predictions when driven by out-of-sample forcing and climate change is an example of this, requiring understanding responses to atmospheric Greenhouse Gas (GHG) concentrations that may be substantially higher than present or the recent past. Notable, though, is the emerging AI technique of “equation discovery”. AI-derived equations from data also does not automatically guarantee good performance for new forcing regimes. However, access to equations rather than a statistical emulator guides system understanding, as their variables and parameters often have a better interpretation. Better process knowledge enables judgements as to whether equations are trusted under extrapolation. For many climate system attributes, descriptive equations are not yet fully available or may be unreliable. This uncertainty is hindering the development of Earth System Models (ESMs) which remain the main tool for projections of large-scale environmental change as GHGs rise. Here, we make the case for using AI-driven equation discovery in climate research, given that its outputs are more interpretable in terms of processes. As ESMs are based around the numerical discretisation of equations that describe climate components, equation discovery from new datasets provides a format amenable to direct inclusion into such models where representation of environmental systems is missing. We present three illustrative examples of how AI-led equation discovery may advance future climate science research. These are generating new equations related to atmospheric convection, parameter derivation for existing equations of the terrestrial carbon cycle, and (additional to ESM improvement) the creation of simplified models of large-scale oceanic features to assess Tipping Point (TP) risks.
- Preprint
(2720 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 23 Oct 2024)
-
CC1: 'Symbolic Regression, Cross-Validation, Examples comment on esd-2024-30', Paul Pukite, 11 Sep 2024
reply
See the attached review (click on lower right link IIRC) .
Overall, the paper is a worthwhile read
-
RC1: 'Comment on esd-2024-30', Anonymous Referee #1, 27 Sep 2024
reply
Overall, I find this submission to be a relevant contribution to the climate science community as it provides a comprehensive overview of ongoing research, outlines current road-blocks and specifically suggests promising approaches which were previously over-looked or unknown in the field. Therefore, I highly suggest to improve upon the clarity and structure of the abstract, introduction and conclusion. This is necessary, as I find the importance of this contribution sometimes gets lost in overly long and disjunct paragraphs and sentences. While, I like the designated re-iteration of the potential application examples in the conclusions, I suggest to improve these by focussing on the discussed ways of using equation discovery in each application. My detailed review can be found in the supplement.
-
RC2: 'Comment on esd-2024-30', Sebastian Scher, 03 Oct 2024
reply
Review ESD
Overall:
The topic of the paper is timely and highly interesting, and the ideas presented in the paper are definitely worth publishing. The introduction part and the method part are well written and give a good overview. The main issue I see with the paper in its current form is that while the intention of the paper is obviously to provide ideas for AI solutions, most of the example part (part 3) of the paper discusses existing solutions or problem settings, without reference to potentially new solutions. For example, all three figures that describe the examples do not make any reference to any proposed new solutions.
Having said that, the ideas are definitely there in the paper, but especially in part 3, they tend to drown in overly long discussions of the domains of the examples, instead on focusing on potential use of AI.
Therefore, while there is nothing fundamentally wrong or flawed with the paper in its current form, I would strongly encourage the authors to restructure especially part 3. Some ideas:
- Make the text focus more on potential AI solutions
- Replace the existing figures – which have a lot of details on the domains – with conceptual figures. With this, readers would be at a single glance be able to spot your ideas
Other comments
it might be worthy to briefly mention explainable AI as well, and what differentiates equation discovery from it.
p.5 maybe mention weather clustering, which is widely used
part 3.1: it is a bit unclear to what the task of an AI part would be in detail. There is in my opinion too much detail of the physics around it, and too little detail on how to solve the problems using equation discovery. E.g., instead of fig.2, which shows a lot of details on a convective storm (something that is actually not the topic of this paper), I think it would be more beneficial to have a conceptual graphic showing how to use equation discovery in this context.
3.2: same as for the first example: a conceptual figure on what is actually attempted would be very valuable.
p.15 L5-6: ” ML-derived spatial aggregation is a form of technique known as computer vision”. Computer vision is a very broad field, so this sentence is incorrect. I also do not understand what you exactly mean here with “ML-derived”. Again, less focus on the state of the field (here carbon cycle modeling) and more focus on the actual AI ideas would be better.
3.3 in this section, even more than in 3.1 and 3.2, there is too much focus on existing methods and models, and the ideas using AI get lost in it. For example, I do not see the need of the detailed discussion, including equations, of the simple ENSO model.
p.15 L 22-23 “However, the large computational time of such simulations maintains interest in faster summary models, mainly in the form of coupled Ordinary Differential Equations (ODEs).” This should be reformulated, as it is a bit confusing, since also the high resolution ESMs are based on coupled ODEs.
p.22 L3: the word “threat” seems an odd choice here.
In the paper, it is mentioned a couple of times (e.g. p21 L26) that it is unclear whether AI-developed models, because of their statistical nature, can extrapolate beyond current forcings. In our paper on Lorenz models, we showed – albeit in a highly simplified setting – that AI models indeed can to some extent learn the influence of external forcing and extrapolate it (https://npg.copernicus.org/articles/26/381/2019/npg-26-381-2019.html). This does of course not necessarily generalize to more complex models, but it might be worth mentioning.
General style:
Abbreviations: this is clearly a matter of personal taste, but considering the interdisciplinary nature of this paper, I would suggest to reduce the use of abbreviations. For example, TP (Tipping Points) is an abbreviation that many readers might not be familiar with, and it is anyway used used only a couple of times. Therefore, it would make the paper easier to read if it simply spelled out every single time (when I encountered it in the conclusion section, it took me some time to remember it meant tipping point even though it was mentioned in the introduction).
Citation: https://doi.org/10.5194/esd-2024-30-RC2
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
305 | 61 | 41 | 407 | 5 | 7 |
- HTML: 305
- PDF: 61
- XML: 41
- Total: 407
- BibTeX: 5
- EndNote: 7
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1