Accurately predicting regional climate change under various emission scenarios is crucial for effective mitigation and adaptation strategies. However, the computational cost of state-of-the-art Global Climate Models (GCMs) limits the exploration of diverse emission pathways and their consequences. This study proposes using machine learning to create a faster, surrogate model that predicts long-term climate responses based on shorter simulations. This is motivated by existing research suggesting links between short-term and long-term climate response patterns. The Hadley Centre Global Environment Model 3 (HadGEM3) provides a unique dataset of simulations with stepwise perturbations to various forcing agents, allowing for the training of machine learning models. The goal is to learn a function that maps short-term (first 10 years) temperature responses to long-term (years 70-100) responses for each grid cell, considering both global and regional forcing scenarios. The study utilizes Ridge regression and Gaussian Process Regression (GPR) due to their ability to handle the limited sample size of available simulations, offering a more efficient alternative to computationally expensive GCM runs for long-term climate projections.

The paper reviews existing literature on climate change modeling, highlighting the computational limitations of GCMs and the need for faster projection methods. It cites studies suggesting relationships between short-term and long-term climate responses, which form the basis for the proposed machine learning approach. The authors note that while data science methods are increasingly used in climate science, no previous study has attempted to predict long-term climate responses across a wide range of forcing scenarios using this specific method. The review acknowledges the use of pattern scaling methods, which the paper then uses as a benchmark for comparison with the proposed machine learning approaches.

The study uses a unique dataset of HadGEM3 simulations, including perturbations to long-lived greenhouse gases (CO2, CH4), short-lived pollutants (sulfate, black carbon), and solar forcing. The simulations provide both short-term (first 10 years) and long-term (years 70-100) surface temperature responses. Two machine learning methods are employed: Ridge regression and Gaussian Process Regression (GPR). Both methods are applied independently to each grid cell, allowing predictions to utilize information from global temperature changes. Ridge regression introduces a regularization constraint to handle the high dimensionality of the data (27,840 grid cells) and limited sample size (21 simulations). GPR, a non-parametric Bayesian approach, models a distribution over possible functions mapping short-term to long-term responses. Model performance is evaluated using root mean squared error (RMSE) at the grid-cell level and absolute error in regional and global mean temperatures, comparing predictions to the HadGEM3 simulations. The study also benchmarks the machine learning methods against a traditional pattern scaling approach, which scales the long-term response pattern of a 2xCO2 scenario based on the ratio of effective radiative forcing (ERF). Cross-validation is employed to optimize the regularization parameter in Ridge regression and hyperparameters in GPR. The study explores the influence of different variables, time periods, and dimension reduction techniques on prediction accuracy. Alternative pattern scaling approaches and other machine learning algorithms (LASSO and Random Forest) are also considered but ultimately found to be less effective.

Both Ridge regression and GPR outperform the pattern scaling approach in predicting long-term climate responses, especially for short-lived pollutant scenarios. The machine learning models better capture regional patterns and diversity, particularly the specific regional impacts of aerosol forcings, which are underestimated by pattern scaling. At the grid-scale level, both machine learning methods capture broad features such as enhanced warming over the Northern Hemisphere and Arctic amplification. However, grid-scale RMSE values need cautious interpretation due to potential bias from small spatial offsets. Analyzing absolute errors at larger spatial scales (global mean temperature and regional means across ten major world regions) shows consistent superior performance of both machine learning models over pattern scaling, with GPR often yielding the lowest errors. The analysis of Ridge regression coefficients reveals patterns in short-term responses that indicate long-term temperature changes, highlighting that predictability varies regionally and involves complex interactions between distant regions. Averaging the magnitude of coefficients across all grid cells shows amplified weights in sea-ice regions, high-altitude regions, major emission areas, and mid-latitude jet stream regions. Increasing the number of training simulations significantly improves prediction accuracy, particularly in regions such as Europe, where the response is highly variable and more data are needed to reduce uncertainty. Europe is identified as a particularly challenging region due to large variations in long-term responses and higher internal variability, indicating a weaker signal-to-noise ratio for the models.

The findings demonstrate the potential of machine learning to significantly improve the efficiency and accuracy of long-term climate change projections. The ability to predict long-term climate responses from shorter simulations drastically reduces the computational cost of climate modeling, enabling more comprehensive scenario explorations. The superior performance of machine learning, especially for regional patterns and short-lived pollutant scenarios, highlights the limitations of traditional pattern scaling methods. The identification of early indicators of long-term climate change through the analysis of regression coefficients is valuable for climate change detection and attribution studies. However, the study also acknowledges the limitations of using a single GCM (HadGEM3) and emphasizes the need for broader data availability to improve the robustness and generalizability of the machine learning models.

This study successfully demonstrates the application of machine learning to predict long-term climate responses from short-term GCM simulations. The machine learning models (Ridge regression and GPR) outperform traditional pattern scaling approaches, particularly for capturing regional variations in response, especially for short-lived pollutants. The study highlights the importance of larger, more diverse datasets for improving model accuracy and suggests focusing on increasing the number of simulations, especially those of short-lived pollutants. Future research should focus on expanding the dataset through international data sharing and exploring the applicability of these methods to other climate variables such as precipitation. The approach presents a promising avenue for faster, more efficient, and detailed climate change projections.

The study's reliance on a single GCM (HadGEM3) limits the generalizability of the findings. The limited number of simulations in the training dataset also affects the model's ability to capture the full complexity of climate responses. Specific regions, like Europe, exhibit higher internal variability, making predictions more challenging and underscoring the need for more data in these areas. The use of multidecadal averages to define short- and long-term responses aims to minimize the influence of internal variability but may mask some important temporal dynamics. Furthermore, the focus on surface temperature as the predictor variable may not fully capture the complexity of the climate system. Finally, the methods presented are evaluated based on a dataset of step-wise perturbations, which might be less representative of real world scenarios.