New research could lead to faster, trustworthy AI models for climate prediction. Really!

When climate models are developed to simulate human activities! This is not science, but resembles alchemy!

And the once prestious PNAS even published this article!

Notice also here again the prejudice and bias involved. We still know barely anything about the very complex natural phenomenon of climate, but the MIT is focusing here human activity in form of CO2 (see also Fig. 2 below copied from their article).

What do they mean here by "method to optimize training datasets"? Data manipulation?

"Now, MIT researchers have developed a technique that streamlines the production of fast, trustworthy [???], and adaptable artificial intelligence models for climate prediction.

These specialized AI models, known as climate emulators, can rapidly simulate the effects of human activities [???] onto a future climate, predicting impacts on temperatures or sea level, for example. ..."

From the significance and abstract:

"Significance

Climate emulators developed through machine learning promise to revolutionize impact assessments by offering an inexpensive way to explore future climates. However, these models often fail to generalize to scenarios that differ from their training data.

While current research emphasizes improving model designs, we demonstrate that the training data themselves can be a limiting factor. We introduce a method to optimize training datasets, such that emulators demonstrate improved generalization to unseen scenarios. Our results suggest an opportunity for modeling centers to reallocate resources to simulate scenarios tailored for training emulators, thereby enabling future climate pathways to be explored at a fraction of the cost.

Abstract

As deep learning for physical systems continues to grow in popularity, efforts to improve generalizability have primarily focused on designing architectures that embed physical constraints. However, for machine-learning surrogate climate models (emulators), we show that the low structural diversity in existing scenarios commonly used to generate training data places a ceiling on predictive skill.

Here, we examine whether training datasets themselves can be optimized to improve generalization.

We introduce a method to create datasets that produce emulators capable of generalizing to new, structurally different scenarios absent from the training data.

We use a differentiable simple climate model (SCM) to calculate the sensitivity of emulator loss to perturbations in the training data, iteratively updating the training data to maximize emulator skill. For an SCM, an emulator trained on one scenario optimized in this fashion outperforms an emulator trained on six standard Scenario Model Intercomparison Project (ScenarioMIP) pathways. We achieve this higher predictive skill despite training on a smaller dataset, finding that our emulator successfully isolates distinct physical behaviors of different climate forcing agents (e.g., greenhouse gases vs. aerosols) without single-forcing runs.

We then demonstrate that scenarios optimized using an SCM, when used to drive an intermediate-complexity climate model, produce a training dataset that yields a more skillful emulator than training on ScenarioMIP outputs.

Our results suggest that, in the compute-constrained environment of running full-scale climate models, generating a small number of dynamically rich scenarios provides greater marginal value for emulation and characterizing system responses than expanding the suite of traditional emissions pathways."

New research could lead to faster, trustworthy AI models for climate prediction | MIT News | Massachusetts Institute of Technology "MIT scientists developed a training method to generate models that simulate the effects of human activities [ ???] on climate much more quickly than other approaches."

Optimal scenario design for climate emulation (no public access)

Optimal scenario design for climate emulation (preprint, open access)