A climate science breakthrough likely won’t be on some massive computation but on a novel formulation that exposes some fundamental pattern (perhaps discovered by deep mining during a machine learning exercise). Over 10 years ago, I wrote on a blog post on how one can extract the ENSO signal by doing simple signal processing on a sea-level height (SLH) tidal time-series — in this case, at Fort Denison located in Sydney harbor.
The formulation/trick is to take the difference between the SLH reading and that from 2 years (24 months) prior, described here
Check the recent blog post Lunar Torque Controls All for context of how it fits in to the unified model.
The rationale for this 24 month difference is likely related to the sloshing of the ocean triggered on an annual basis. I think this is a pattern that any ML exercise would find with very little effort. After all, it didn’t take me that long to find it. But the point is that the ML configuration has to be open and flexible enough to be able to search, generate, and test for the same formulation. IOW, it may not find it if the configuration, perhaps focused on computationally massive PDEs, is too narrow. That was my comment to a RC post on applying machine learning to climate science, see the following link and subsequent quote:
Nick McGreivy commented on:
“ML-based parameterizations have to work well for thousands of years of simulations, and thus need to be very stable (no random glitches or periodic blow-ups) (harder than you might think). Bias corrections based on historical observations might not generalize correctly in the future.”
This same issue arises when using ML to simulate PDEs. The solution is to analytically calculate what the stability condition(s) is (are), then at each timestep to add some numerical diffusion that nudges the solution towards satisfying the stability condition(s). I imagine this same technique could be used for ML-based parametrizations.
Elsewhere in that RC thread https://www.realclimate.org/index.php/archives/2024/12/ai-caramba/#comment-828559
—-
What climate scientists trying to use NN haven’t learned yet is the closed-world assumption (CWA) that is at the core of classic AI. Neural networks are trained on a fixed dataset and if this is embodied only by the climate data itself, it will never be aware of information outside this dataset. That’s the closed world aspect, and unless all the relevant inputs are included, which are essentially the guiding boundary-values, they will likely be spinning their wheels.. Whatever gets produced will be an unresolved convolution of some unknown forcing with an also unknown response function — in other words, the fitted model is still totally encrypted! The NN essentially fits patterns without truly disentangling causation, so still no way to decode the resultant fit with meaningful insight, and thus highly unlikely to be of any use for predictive extrapolations..
I scan many of these machine learning climate papers and spend little effort if the authors do not acknowledge their closed-world assumptions.
It doesn’t have to be a statistical fitting. One can also generate a deterministic model from a ML training exercise. That’s essentially the same thing as a regression yielding the transfer function from an input. If you don’t think this aspect is important, consider autonomous driving — make the statistical or probability uncertainty window too big and expect many crashes. Many deterministic constraints involved in an autonomous situation. My favorite related example is tidal analysis — it is almost strictly deterministic and the remaining uncertainty is more than likely unresolved tidal factors.
This will be an interesting climate topic for years to come.
—-
This is what machine learning experiments will be finding, tiny effects that people give up on..
Patrick in https://www.realclimate.org/index.php/archives/2024/11/unforced-variations-dec-2024/#comment-828641 said:
Whether an effect is tiny or not is a matter of scale. In the greater scheme of things, like dLOD, the QBO itself is pretty insignificant — as it’s a thin band of very low density atmosphere wrapped around the equator. Not much there really, but science is not always about big vs tiny effects. After all, CO2 is tiny too.
Yes, if the moon stayed in the same plane as the ecliptic orbit, it’s torque would be aligned with the sun’s torque. The angular momentum vector L of the degenerate tropical/draconic orbit would point the same way as the ecliptic orbit (orthogonal to the ecliptic plane), so at most would modulate the strength of the annual cycle. Thus, there would not be a different angular momentum vector L1 that would cause another wobble (i.e. Chandler) to beat with the annual wobble, or non-congruent vector torque to compete with the semi-annual oscillation (SAO) and thus form a QBO. That’s all part of the group symmetry argument I am offering.
Very few geophysicists want to take this on, as it overturns decades of conventional wisdom. I would not even be considering it if the numbers for the Chandler wobble and QBO didn’t match exactly to this model. That plus I have a strong inkling that the massive amounts of machine learning applied by the big guns will eventually cover this same ground and I want to be able to be ready for that. ML experiments search for numerical patterns and do extensive cross-validation to avoid over-fitting so climate scientists should take heed in case they “discover” the same agreement.
Pingback: Difference Model Fitting | GeoEnergy Math