A key consideration for evaluating climate projections is uncertainty in radiative forcing scenarios. Although it is straightforward to monitor greenhouse gas concentrations and compare those observations with specified climate scenarios, it remains less obvious on how to connect regional climate patterns with these scenarios in real time. Here we introduce a machine learning approach for linking patterns of climate change with radiative forcing scenarios and use an attribution method to understand how these linkages are made. We train a neural network using output from the SPEAR Large Ensemble to classify whether temperature or precipitation maps are most likely to originate from one of several potential radiative forcing scenarios. The neural network learns to identify “fingerprint” patterns that associate signals of climate change with the scenarios. We illustrate this using output from additional mitigation experiments and highlight regions that are critical for associating the new simulations with likely radiative forcing scenarios.

To better understand the regional changes in summertime temperatures across the conterminous United States (CONUS), we adopt a recently developed machine learning framework that can be used to reveal the timing of emergence of forced climate signals from the noise of internal climate variability. Specifically, we train an artificial neural network (ANN) on seasonally-averaged temperatures across the CONUS and then task the ANN to output the year associated with an individual map. In order to correctly identify the year, the ANN must therefore learn time-evolving patterns of climate change amidst the noise of internal climate variability. The ANNs are first trained and tested on data from large ensembles and then evaluated using observations from a station-based dataset. To understand how the ANN is making its predictions, we leverage a collection of ad hoc feature attribution methods from explainable artificial intelligence (XAI). We find that anthropogenic signals in seasonal mean minimum temperature have emerged by the early 2000s for the CONUS, which occurred earliest in the Eastern United States. While our observational timing of emergence estimates are not as sensitive to the spatial resolution of the training data, we find a notable improvement in ANN skill using a higher resolution climate model, especially for its early 20th century predictions. Composites of XAI maps reveal that this improvement is linked to temperatures around higher topography. We find that increases in spatial resolution of the ANN training data may yield benefits for machine learning applications in climate science.

Evaluating historical simulations from global climate models (GCMs) remains an important exercise for better understanding future projections of climate change and variability in rapidly warming regions, such as the Arctic. As an alternative approach for comparing climate models and observations, we set up a machine learning classification task using a shallow artificial neural network (ANN). Specifically, we train an ANN on maps of annual mean near-surface temperature in the Arctic from a multi-model large ensemble archive in order to classify which GCM produced each temperature map. After training our ANN on data from the large ensembles, we input annual mean maps of Arctic temperature from observational reanalysis and sort the prediction output according to increasing values of the ANN’s confidence for each GCM class. To attempt to understand how the ANN is classifying each temperature map with a GCM, we leverage a feature attribution method from explainable artificial intelligence. By comparing composites from the attribution method for every GCM classification, we find that the ANN is learning regional temperature patterns in the Arctic that are unique to each GCM relative to the multi-model mean ensemble. In agreement with recent studies, we show that ANNs can be useful tools for extracting regional climate signals in GCMs and observations.

The global mean surface temperature (GMST) record exhibits both interannual to multidecadal variability and a long-term warming trend due to external climate forcing. To explore the predictability of temporary slowdowns in decadal warming, we apply an artificial neural network (ANN) to climate model data from the Community Earth System Model Version 2 Large Ensemble. Here, an ANN is tasked with whether or not there will be a slowdown in the rate of the GMST trend by using maps of ocean heat content at the onset. Through a machine learning explainability method, we find the ANN is learning off-equatorial patterns of anomalous ocean heat content that resemble transitions in the phase of the Interdecadal Pacific Oscillation in order to make slowdown predictions. Finally, we test our ANN on observed historical data, which further reveals how explainable neural networks are useful tools for understanding decadal variability in both climate models and observations.

It remains difficult to disentangle the relative influences of aerosols and greenhouse gases on regional surface temperature trends in the context of global climate change. To address this issue, we use a new collection of initial-condition large ensembles from the Community Earth System Model version 1 that are prescribed with different combinations of industrial aerosol and greenhouse gas forcing. To compare the climate response to these external forcings, we adopt an artificial neural network (ANN) architecture from previous work that predicts the year by training on maps of near-surface temperature. We then utilize layer-wise relevance propagation (LRP) to visualize the regional temperature signals that are important for the ANN’s prediction in each climate model experiment. To mask noise when extracting only the most robust climate patterns from LRP, we introduce a simple uncertainty metric that can be adopted to other explainable artificial intelligence (AI) problems. We find that the North Atlantic, Southern Ocean, and Southeast Asia are key regions of importance for the neural network to make its prediction, especially prior to the early-21st century. Notably, we also find that the ANN predictions based on maps of observations correlate higher to the actual year after training on the large ensemble experiment with industrial aerosols held fixed to 1920 levels. This work illustrates the sensitivity of regional temperature signals to changes in aerosol forcing in historical simulations. By using explainable AI methods, we have the opportunity to improve our understanding of (non)linear combinations of anthropogenic forcings in state-of-the-art global climate models.