Using neural networks to explore regional climate patterns in single-forcing large ensembles
Abstract
Understanding the time-evolving influence of aerosols and greenhouse gases on regional climate variability remains challenging for climate attribution and detection. To investigate the climate response to these complex interactions, we use a collection of single-forcing large ensembles from the Community Earth System Model version 1, which are prescribed with different combinations of industrial aerosol and greenhouse gas forcing. To disentangle the unique climate signals in each large ensemble simulation, we use an artificial neural network (ANN) that outputs the year for each global map of near-surface temperature. We then employ an explainable artificial intelligence (XAI) method to compare what temperature patterns the ANN is using to make each prediction. In particular, we find that the North Atlantic Ocean is an important region for the ANN to make its prediction, especially for the large ensemble simulation with industrial aerosols held fixed to 1920 levels. The Southern Ocean and Southeast Asia are also regions of reliable climate signals that the ANNs leverages to predict the year for each temperature map. Lastly, we test our XAI framework on observations using temperature data from 20th century reanalysis. We find that XAI methods can be used to better understand patterns of forced climate change on regional to global scales. This work further emphasizes the importance of understanding the sensitivity of global climate models to changes in prescribed aerosol forcing.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.A52E..01L