Climate Intervention Analysis using AI Model Guided by Statistical Physics Principles

Machine learning (ML) offers significant potential for accelerating scientific discovery, particularly in domains with computationally expensive simulations. Estimating a system's response to external forcings, however, typically requires extensive labeled data or costly simulations. This is especially critical in climate science, where Earth System Models (ESMs) are used to understand climate change and potential climate interventions. ESMs provide comprehensive simulations of Earth's complex systems, but a single run can take weeks or months. Climate intervention strategies, such as Solar Radiation Management (SRM), including Stratospheric Aerosol Injection (SAI) and Marine Cloud Brightening (MCB), aim to mitigate climate change by manipulating the Earth's environment. MCB, involving the spraying of sea salt aerosols into low-level clouds, presents a particularly challenging case due to its regional, spatially heterogeneous effects and the vast number of potential intervention scenarios. Existing evaluations of MCB have been limited by computational constraints. While ML has shown promise in various climate modeling contexts, predicting the outcome of external forcings remains difficult due to limited training data. This study addresses this challenge by proposing a novel approach using the Fluctuation-Dissipation Theorem (FDT) from statistical physics to guide the development of an AI model for rapid climate intervention analysis.

The application of machine learning in scientific disciplines has yielded significant advancements in various areas, including numerical weather prediction, medical image diagnosis, astronomical image classification, and drug discovery. These applications often involve direct application of ML algorithms to large datasets or utilize surrogate models trained on simulation data to efficiently reproduce model outputs. However, predicting a system's response to external forcings, a crucial aspect of systems analysis and engineering, remains a significant challenge due to the computational costs associated with running extensive simulations. The Fluctuation-Dissipation Theorem (FDT), a fundamental principle in statistical physics, relates a system's response to external forces to its internal fluctuations. Previous work has applied FDT in ML to ensure physical consistency in learned models, such as thermodynamic consistency in learning reversible and irreversible dynamics. This study, however, takes a novel approach by applying FDT to learn from internal climate variability to estimate the forced response. This approach differs from previous works which focus primarily on enforcing physical constraints rather than utilizing internal variability for prediction.

The proposed AiBEDO framework uses a two-phase approach to rapidly generate MCB intervention impact assessments. Phase 1 involves creating an AI emulator to map the relationships between input (cloud and clear-sky radiation anomalies at time t, x(t)) and output (surface climate anomalies after a time delay τ, y(t+τ)) variables. Three machine learning models – Spherical Multilayer Perceptron (S-MLP), Spherical U-Net (S-Unet), and Spherical Adaptive Fourier Neural Operator (S-AFNO) – were trained to learn the mapping Aτ: x(t) → y(t+τ) for different time lags (τ). The models were trained using a large ensemble of Earth System Model (CESM2-LE) simulations comprising over 100,000 model months. To address the non-uniform area of typical latitude-longitude grids, the data underwent a geodesy-aware spherical sampling technique, transforming it into a 1D spherical icosahedral mesh using PyGSP library. Data preprocessing involved subtracting the ensemble mean to remove seasonal cycles and secular trends, focusing on internal climate variability. Phase 2 involves summing the time-lagged mappings from Phase 1 using an FDT-like operator to obtain a time-integrated estimate of the regional impact of the external forcing. This approach accounts for non-linear components of the climate response, thereby relaxing the Gaussianity assumption of traditional FDT applications. This time-integrated output, denoted by δy(t), is calculated by summing the differences between AiBEDO outputs for perturbed and unperturbed internal fluctuations across various time lags. To evaluate AiBEDO's accuracy, the model's predictions were compared against targeted ESM runs with MCB-like forcings. The comparison focused on three regions: Northeast Pacific (NEP), Southeast Pacific (SEP), and Southeast Atlantic (SEA). Finally, an interactive visualization platform was developed to facilitate post-hoc analysis and rapid prototyping of MCB scenarios.

The comparative analysis of the three ML models (S-MLP, S-Unet, and S-AFNO) revealed that S-MLP consistently outperformed the others in emulating the CESM2-LE data across different time lags. While S-AFNO showed comparable performance to S-MLP for simultaneous predictions, S-MLP demonstrated superior accuracy in capturing global response patterns and long-range interactions, especially in time-lagged scenarios. The S-Unet's lower performance is attributed to the loss of information regarding global climate patterns during spherical convolution and pooling. The MCB experiments using the superior S-MLP model within AiBEDO showed high fidelity in reproducing the climate response patterns observed in the targeted CESM2 simulations, achieving correlation scores of 0.68 for temperature (tas), 0.51 for precipitation (pr), and 0.47 for pressure (ps). AiBEDO effectively captured remote teleconnected responses and correctly identified climate responses to different forcing regions in several key areas. However, AiBEDO's performance was weaker for regional perturbations compared to the simultaneous perturbation of all three regions, and it performed better in the tropics and over oceans compared to higher latitudes and over land. The interactive visualization platform successfully enabled rapid prototyping of MCB scenarios, providing insights into regional climate impacts and potential risks to climate tipping points. This platform allows users to specify regions, input variables, and perturbation extent, visualize results, and assess risks to seven identified regional tipping points.

The results demonstrate the feasibility of using FDT-guided AI models to significantly accelerate climate intervention analysis. The AiBEDO framework, particularly using the S-MLP model, achieves high-fidelity emulation of ESM simulations, offering a three-order-of-magnitude speedup. The model's ability to capture remote teleconnected responses is crucial for assessing the potential unintended consequences of MCB. The interactive visualization platform enhances the usability and interpretability of AiBEDO, facilitating informed decision-making in climate intervention strategies. The study's focus on MCB provides a powerful example of the framework's applicability, but the underlying methodology is transferable to various computationally demanding scientific domains. This approach helps bridge the gap between complex simulation models and the need for rapid scenario analysis.

This study introduces AiBEDO, a novel framework for climate intervention analysis that leverages the Fluctuation-Dissipation Theorem and machine learning. AiBEDO, particularly the S-MLP implementation, efficiently emulates complex ESM simulations, providing high-fidelity predictions at significantly increased speeds. The associated visualization platform facilitates rapid prototyping and informed decision-making for climate intervention strategies. Future work will focus on optimizing MCB perturbations to mitigate specific climate tipping points, potentially incorporating constraints to avoid unintended consequences. This will expand AiBEDO's capabilities as a valuable tool for climate science and policy.

The study's primary limitation lies in the reliance on CESM2-LE data, which may limit the generalizability of findings to other ESMs. The model's performance at high latitudes and over land is weaker than in tropical ocean regions. Furthermore, while the model captures many aspects of the climate response, it might still miss some subtle or complex interactions that would be revealed by more extensive ESM simulations. The selection of seven climate tipping points was based on existing literature and might not encompass all relevant tipping points.