AI Trained on Climate Models Predicts Warmer Faster
Essay by Eric Worrall
Why did they train the AI with climate models in the first place? Why not ignore the models and directly use the observations to train the AIs directly?
AI study predicts planet will warm faster than expected
According to CNN 11:27 a.m. January 31, 2023
…
The study estimated that the planet could reach 1.5 degrees Celsius of warming above pre-industrial levels in a decade and found there was a “significant possibility” of global temperature rise exceeding 2 degrees Celsius. mid-century, even as there are significant efforts globally to bring about reductions in planet-warming pollution.
The data shows that the average global temperature has increased by about 1.1 degrees to 1.2 degrees since industrialization.
“Our results provide further evidence of climate change having a major impact over the next three decades,” notes reportpublished Monday in the journal Proceedings of the National Academy of Sciences.
…
Summary of the study;
Data-driven prediction of time remaining until critical global warming threshold is reached
Noah S. Diffenbaugh and Elizabeth A. Barnes
Michael Mann, Pennsylvania State University, University Park, PA; received April 25, 2022; accept December 14, 2022
January 30, 2023
120 (6) e2207183120
meaning
The UN Paris Agreement aims to keep global warming below 2°C and pursues the 1.5°C target. Given the clear evidence of increasing climate impacts, the time remaining until these global thresholds are reached is a topic of considerable interest. We use machine learning methods to make truly out of the box predictions about that time, based on spatial modeling of historical temperature observations. Our results confirm that global warming is on the verge of surpassing the 1.5 °C threshold, even if the course of climate impacts declines significantly in the near term. Our projections also suggest that even with significant greenhouse gas reductions, it is unlikely to keep global warming below the 2 °C threshold.
abstract
Leveraging artificial neural networks (ANNs) trained on the output of a climate model, we use a spatial model of historical temperature observations to predict the time until reaching important global warming threshold. Although no observations were used during training, validation, or testing, ANN accurately predicted the historical moment of global warming from historical annual temperature maps. The central estimate for the 1.5°C global warming threshold is from 2033 to 2035, covering a range of ±1σ from 2028 to 2039 in the Intermediate Climate Forced scenario (SSP2-4.5), consistent with previous reviews. However, our data-driven approach also shows a significant possibility of crossing the 2 °C threshold even in the Low climate imperative scenario (SSP1-2.6). Despite the limitations of our approach, our results suggest a higher likelihood of reaching 2 °C in the Low scenario than indicated in some previous assessments—although does not exclude the possibility that 2 °C can be avoided. Interpretable AI methods reveal that ANNs focus on specific geographic regions to predict the time until a global threshold is reached. Our framework provides a unique, data-driven approach to quantifying the signal of climate change in historical observations and to limit uncertainty in climate model projections. Given the substantial evidence currently available for increased risks to natural and human systems at 1.5 °C and 2 °C, our results provide further evidence that climate change has major impact over the next three decades.
…
Read more: https://www.pnas.org/doi/full/10.1073/pnas.2207183120
For the credit, they have published their code on github. “… The code is available on GitHub at https://github.com/eabarnes1010/target_temp_detection (60) and hosted on Zenodo at the following DOI: https://doi.org/10.5281/zenodo.7510551 (sixty one).”.
I understand that researchers are trying to use AI to identify geographically important areas in terms of climate or the distribution of observations, in an attempt to filter out noise and reduce the uncertainty of predictions.
My concern with this approach is that if the data is enough to adjust the predictions, the AI can be trained directly on the data, the AI can infer climate models directly from the data.
The use of a simulation or model allows to encapsulate a large number of training runs in a short period of time and limit the output of the AI. But the AI is then tainted by the model, which actually becomes an extension of the model.
I guess time will tell if their approach yields increased predictive skills.