From Watts Up With That?

Essay by Eric Worrall

If you have a possible missing variable problem, the solution is to add more arbitrary adjustments to your model?

AI Exposes Accelerated Climate Change: 3°C Temperature Rise Imminent

BY IOP PUBLISHING DECEMBER 28, 2024

AI-enhanced research shows regional warming will exceed critical thresholds faster than expected, with most regions surpassing 1.5°C by 2040. Vulnerable areas like South Asia face heightened risks, urging swift adaptation actions.

Three leading climate scientists have analyzed data from 10 global climate models, utilizing artificial intelligence (AI) to enhance accuracy. Their findings indicate that regional warming thresholds are likely to be reached sooner than previously estimated.

…

Elizabeth Barnes says: “Our research underscores the importance of incorporating innovative AI techniques like transfer learning into climate modeling to potentially improve and constrain regional forecasts and provide actionable insights for policymakers, scientists, and communities worldwide.”

…Read more:

 https://scitechdaily.com/ai-exposes-accelerated-climate-change-3c-temperature-rise-imminent/

The referenced study;

Combining climate models and observations to predict the time remaining until regional warming thresholds are reached

Elizabeth A Barnes^*, Noah S Diffenbaugh and Sonia I Seneviratne

Published 10 December 2024 • © 2024 The Author(s). Published by IOP Publishing Ltd
Environmental Research Letters, Volume 20, Number 1 Citation Elizabeth A Barnes et al 2025 Environ. Res. Lett. 20 014008DOI 10.1088/1748-9326/ad91ca

Abstract

The importance of climate change for driving adverse climate impacts has motivated substantial effort to understand the rate and magnitude of regional climate change in different parts of the world. However, despite decades of research, there is substantial uncertainty in the time remaining until specific regional temperature thresholds are reached, with climate models often disagreeing both on the warming that has occurred to-date, as well as the warming that might be experienced in the next few decades. Here, we adapt a recent machine learning approach to train a convolutional neural network to predict the time (and its uncertainty) until different regional warming thresholds are reached based on the current state of the climate system. In addition to predicting regional rather than global warming thresholds, we include a transfer learning step in which the climate-model-trained network is fine-tuned with limited observations, which further improves predictions of the real world. Using observed 2023 temperature anomalies to define the current climate state, our method yields a central estimate of 2040 or earlier for reaching the 1.5 °C threshold for all regions where transfer learning is possible, and a central estimate of 2040 or earlier for reaching the 2.0 °C threshold for 31 out of 34 regions. For 3.0 °C, 26 °C out of 34 regions are predicted to reach the threshold by 2060. Our results highlight the power of transfer learning as a tool to combine a suite of climate model projections with observations to produce constrained predictions of future temperatures based on the current climate.Read more: 

https://iopscience.iop.org/article/10.1088/1748-9326/ad91ca

If I have understood correctly, they are essentially using the AI as a complex black box polynomial correction to their rather imprecise climate models, to try to squeeze out better answers. The polynomial is trained by comparing observed temperature data to model output, then the resultant amalgamation of climate models and AI polynomial corrections is extrapolated to try to predict future events.

The problem with this approach is it creates the illusion of accuracy, without actually knowing if greater accuracy has been achieved. An AI used in this way applies complex arbitrary “corrections” to input data, to generate a near perfect match to any data used to train that AI. But the AI knows nothing about the underlying physical phenomena. The AI might be able infer physical phenomena if it has enough data – or the AI could just make stuff up, especially if unknown critical input data is missing from the set of data which is used to train the AI.

AI does have a role in scientific analysis. In fields like drug discovery and complex optimisation problems, AI can produce excellent results.

But AI also has a well known tendency to go off the rails, to “hallucinate” false results.

An AI malfunction is not a problem if you can test the quality of the AI results immediately. But using AI to try to figure out how to correct climate models, where nobody will know for years or decades whether the AI got it right, then using those AI corrections to project future events, this seems a dubious use of artificial intelligence.