Comparison of Climate Model Large Ensembles With Observations in the Arctic Using Simple Neural Networks

Abstract

The Arctic is warming at a rate of more than three times as fast as the globally averaged mean surface temperature trend (Druckenmiller et al., 2021). This dramatic warming, otherwise known as Arctic amplification, is accompanied by long-term losses of Arctic sea-ice extent and thickness (

Identifying the regional emergence of climate patterns in the ARISE-SAI-1.5 simulations

“…Attribution describes the contribution of the input features to the overall output. Despite an increasing number of explainable machine learning methods adopted for various climate science applications (e.g., Toms et al, 2020;Sonnewald and Lguensat, 2021;Labe and Barnes, 2022;Molina et al, 2021), we focus on two conceptually simple methods that we refer to as contribution maps. For identifying the significant regions to determine whether a climate map is from the SSP2-4.5 or SAI simulation, we consider contribution maps by multiplying the logistic regression model weights by the input values for every location on the map.…”

Identifying climate models based on their daily output using machine learning

“…The properties of the CNN as a highly nonlinear, deep learning method, however, did not allow a straightforward extraction of the features used to separate the two categories, but a planned detailed investigation will be able to reveal more details in future work. This will be based, for example, on layer-wise relevance propagation, a technique that can reveal the “grid cell relevance” of the temperature maps used as input for the output probabilities of a trained neural network (Bach et al, 2015; Toms et al, 2020; Labe and Barnes, 2022). For a comprehensive overview of such an approach applied to a related problem, including a review of the current literature, see, for example, Labe and Barnes (2022).…”

“…This will be based, for example, on layer-wise relevance propagation, a technique that can reveal the “grid cell relevance” of the temperature maps used as input for the output probabilities of a trained neural network (Bach et al, 2015; Toms et al, 2020; Labe and Barnes, 2022). For a comprehensive overview of such an approach applied to a related problem, including a review of the current literature, see, for example, Labe and Barnes (2022). These techniques investigating the classifiers themselves will be combined with background knowledge about climate models to provide integrated and interpretable insights into the origins of the classification skill.…”

“…Future applications could build on the approaches illustrated here in several ways: They could add to the model evaluation toolbox (Eyring et al, 2019) and could target, for example, the classification of individual model components (e.g., atmospheric or ocean component), model generations (e.g., CMIP6 versus the roughly 10 years older CMIP5), perturbed parameter ensembles, or strains of model development in general. Following recent work from Labe and Barnes (2022), misclassifications could also be used systematically to draw conclusions about model performance. The classification approaches could be used to pinpoint model–model or model–observation differences and similarities. This could be done by analyzing the spatial scales of separability, that is, whether models on regional domains are less separable than globally, and/or whether this may depend on specific regions.…”

“…A range of recent studies have demonstrated the potential of statistical and machine learning in climate science. These include approaches for seasonal prediction (Gibson et al, 2021), to identify forced signals from spatial patterns of temperature, precipitation, and humidity (Barnes et al, 2019; Sippel et al, 2020; de Vries et al, 2023), to explore the role of single forcing agents (Labe and Barnes, 2021), to predict modes of atmosphere–ocean variability (Gordon et al, 2021), and, most recently, to contrast models and observations (Labe and Barnes, 2022). Here, we investigate the potential of such statistical and machine learning classifiers to separate models from observations and from each other on the basis of daily data, and thus in the presence of considerable noise arising from the internal variability on weather timescales.…”