Current global actions to reduce greenhouse gas emissions are very likely to be insufficient to meet the climate targets outlined under the Paris Agreement. This motivates research on possible methods for intervening in the Earth system to minimize climate risk while decarbonization efforts continue. One such hypothetical climate intervention is stratospheric aerosol injection (SAI), where reflective particles would be released into the stratosphere to cool the planet by reducing solar insolation. The climate response to SAI is not well understood, particularly on short-term time horizons frequently used by decision makers and planning practitioners to assess climate information. This knowledge gap limits informed discussion of SAI outside the scientific community. We demonstrate two framings to explore the climate response in the decade after SAI deployment in modeling experiments with parallel SAI and no-SAI simulations. The first framing, which we call a snapshot around deployment, displays change over time within the SAI scenarios and applies to the question “What happens before and after SAI is deployed in the model?” The second framing, the intervention impact, displays the difference between the SAI and no-SAI simulations, corresponding to the question “What is the impact of a given intervention relative to climate change with no intervention?” We apply these framings to annual mean 2-meter temperature, precipitation, and a precipitation extreme in the first two experiments to use large ensembles of Earth system models that comprehensively represent both the SAI injection process and climate response, and connect these results to implications for other climate variables.

Soil moisture influences near-surface air temperature by partitioning downwelling radiation into latent and sensible heat fluxes, through which dry soils generally lead to higher temperatures. The strength of this coupled soil moisture-temperature (SM-T) relationship is not spatially uniform, and numerous methods have been developed to assess SM-T coupling strength across the globe. These methods tend to involve either idealized climate-model experiments or linear statistical methods which cannot fully capture nonlinear SM-T coupling. In this study, we propose a nonlinear machine learning-based approach for analyzing SM-T coupling and apply this method to various mid-latitude regions using historical reanalysis datasets. We first train convolutional neural networks (CNNs) to predict daily maximum near-surface air temperature (TMAX) given daily SM and geopotential height fields. We then use partial dependence analysis to isolate the average sensitivity of each CNN’s TMAX prediction to the SM input under daily atmospheric conditions. The resulting SM-T relationships broadly agree with previous assessments of SM-T coupling strength. Over many regions, we find nonlinear relationships between the CNN’s TMAX prediction and the SM input map. These nonlinearities suggest that the coupled interactions governing SM-T relationships vary under different SM conditions, but these variations are regionally dependent. We also apply this method to test the influence of SM memory on SM-T coupling and find that our results are consistent with previous studies. Although our study focuses specifically on local SM-T coupling, our machine learning-based method can be extended to investigate other coupled interactions within the climate system using observed or model-derived datasets.

We develop and demonstrate a new interpretable deep learning model specifically designed for image analysis in earth system science applications. The neural network is designed to be inherently interpretable, rather than explained via post hoc methods. This is achieved by training the network to identify parts of training images that act as prototypes for correctly classifying unseen images. The new network architecture extends the interpretable prototype architecture of a previous study in computer science to incorporate absolute location. This is useful for earth system science where images are typically the result of physics-based processes, and the information is often geo-located. Although the network is constrained to only learn via similarities to a small number of learned prototypes, it can be trained to exhibit only a minimal reduction in accuracy compared to non-interpretable architectures. We apply the new model to two earth science use cases: a synthetic data set that loosely represents atmospheric high- and low-pressure systems, and atmospheric reanalysis fields to identify the state of tropical convective activity associated with the Madden-Julian oscillation. In both cases, we demonstrate that considering absolute location greatly improves testing accuracies. Furthermore, the network architecture identifies specific historical dates that capture multivariate, prototypical behaviour of tropical climate variability.

Predicting Pacific Decadal Oscillation (PDO) transitions and understanding the associated mechanisms has proven a critical but challenging task in climate science. As a form of decadal variability, the PDO is associated with both large-scale climate shifts and regional climate predictability. We show that artificial neural networks (ANNs) predict PDO persistence and transitions from 12 months onward. Using layer-wise relevance propagation to investigate the ANN predictions, we demonstrate that the ANNs utilize oceanic patterns that have been previously linked to predictable PDO behavior. For PDO transitions, ANNs recognize a build-up of ocean heat content in the off-equatorial western Pacific 12-27 months before a transition occurs. The results support the continued use of ANNs in climate studies where explainability tools can assist in mechanistic understanding of the climate system.

As anthropogenic activities continue to drive increases in extreme events, the fundamental solution of reducing greenhouse gas emissions remains elusive. Thus, there is growing interest in stratospheric aerosol injection (SAI) to offset some of the most dangerous consequences of climate change. If SAI was deployed at a global scale, it would likely be easy to detect by some metrics. However, the detectability of SAI on extreme events might be more difficult, given the presence of natural climate variability. We examine this question in climate model simulations of SAI. Specifically, we train a logistic regression model to predict whether a map of global extremes came from climate simulations with or without SAI. The timing of accurate predictions is a quantification of the time to detection of SAI impacts. We find that regional changes in extreme temperature and precipitation are robustly detected within 1 and 15 years of initial SAI injection, respectively.

Due to the coupled nature of the earth system, precipitation forecast errors at S2S lead times are caused by a combination of errors/biases from the atmosphere, ocean, ice and land across a range of spatial and temporal scales. We show that UFS precipitation errors over the U.S. at Weeks 3-4 can be directly related to biases in simulating tropical dynamics. In particular, the inability of the UFS to realistically simulate the Madden-Julian oscillation (MJO) leads to biases in the teleconnection to North America that produces these errors. When the tropics are nudged to produce an accurate representation of the MJO and other tropical disturbances, U.S. West Coast precipitation biases are substantially reduced. A clustering analysis is used to show that the greatest forecast improvements with nudging occur during warm ENSO events when MJO convection is in the Indian Ocean and about to move into the Maritime Continent. Physical mechanisms that explain the improvement in tropical-extratropical teleconnections during certain MJO and ENSO states will be discussed. We will also present future plans to combine state-of-the-art developments in machine learning with process-based diagnostics of the tropical moisture and moist static energy (MSE) budgets to understand and correct precipitation biases in coupled UFS hindcasts. In particular, we will discuss how model biases and errors in tropical variability (e.g. MJO) and associated teleconnections to midlatitudes lead to errors in U.S. precipitation on S2S timescales, and present methods to reduce these errors via post-processing on a forecast-by-forecast basis.