Problem

Artificial intelligence models are increasingly integrated into Earth system science and operational forecasting, but their opaque nature hinders trust and understanding of their internal decision-making processes. A significant challenge arises because most Explainable AI (XAI) methods were initially developed for classification tasks, making their direct applicability and suitability questionable for the regression problems commonly encountered in geoscientific modeling, which involve predicting continuous variables and require adherence to physical consistency. This discrepancy creates a critical gap in assessing AI model reliability and ensuring their responsible deployment in high-stakes contexts.

Method

To address these challenges, the authors systematically review and categorize prominent XAI techniques, delving into their theoretical underpinnings. They then apply a carefully selected subset of these methods, including Saliency, InputxGradient, Integrated Gradients, SHAP, and various Layer-wise Relevance Propagation (LRP) rules, to a neural network designed to emulate the Lorenz-63 chaotic system. This well-understood, low-dimensional system serves as a controlled environment to rigorously test and identify the limitations and potential failure modes of general-purpose XAI when applied to regression. Following this experimental analysis, the paper comprehensively surveys existing applications of diverse XAI methods within Earth system sciences, specifically focusing on regression use cases. Finally, the authors contextualize XAI within a typical model development lifecycle, linking specific methodological choices to the distinct needs and concerns of different stakeholder groups involved in operational Earth system science.

Results

The application of XAI methods to the Lorenz-63 system provided crucial insights, demonstrating that the choice of baseline significantly influences attribution results for methods like Integrated Gradients and SHAP, leading to varied explanations. Techniques such as InputxGradient were observed to suppress importance for near-zero input values, potentially masking dynamically relevant information. Similarly, the diverse rules within LRP produced distinct attribution patterns, underscoring the necessity of a well-justified rule selection for regression. The review of XAI in Earth sciences revealed SHAP's widespread use, yet highlighted its difficulties with highly correlated geoscientific inputs. Gradient-based methods, offering parallels to established adjoint sensitivity analysis, demonstrated considerable promise due to their physical interpretability and scalability. The study concludes that current XAI methods only partially satisfy stakeholder requirements across the model development lifecycle, providing diagnostic value but often lacking comprehensive or systematic explanations, necessitating expert judgment to discern physically meaningful signals from methodological artifacts.

Implications

The paper identifies critical areas for improvement in XAI for Earth system regression. There is a pressing need to refine and adapt existing XAI methods to better handle challenges such as appropriate baseline selection, effectively accounting for correlated inputs, and developing robust evaluation benchmarks tailored for regression tasks. Scalability is another major concern, demanding XAI methods that can operate reliably on high-dimensional models. Furthermore, XAI must evolve to integrate diagnostic information seamlessly into AI model workflows, thereby fostering trust, enhancing transparency, and advancing scientific insight and model development. The authors strongly advise against directly transferring classification-centric methods to regression. They emphasize the importance of intentional baseline choices, prioritizing physically plausible inputs, providing justification for aggregating explanation metrics, and critically, aligning explanations with specific stakeholder questions rather than attempting a holistic model explanation, to ensure XAI delivers actionable and context-relevant insights for Earth system science.