Finding the right XAI method -- A Guide for the Evaluation and Ranking of Explainable AI Methods in Climate Science

There is no one-size-fits-all explanation method: The paper finds that different XAI methods have different strengths and weaknesses depending on the task and neural network architecture used. For climate science applications, methods like Integrated Gradients and layer-wise relevance propagation excel at robustness and faithfulness, while sensitivity methods like gradient-based approaches prioritize different qualities—meaning practitioners must choose methods based on their specific needs rather than assuming any single approach works best.

Evaluating AI explanations requires multiple quality criteria: Rather than judging explanations on a single dimension, the researchers identified five key properties to assess: robustness (consistency of explanations), faithfulness (whether explanations truly reflect what the model learned), randomization (avoiding spurious patterns), complexity (interpretability), and localization (precision of highlighted regions). This multi-criteria framework helps ensure AI systems provide genuinely useful explanations rather than just appearing transparent.

Trade-offs exist between explanation properties: The study reveals that XAI methods cannot excel at all qualities simultaneously—for example, methods strong in robustness and faithfulness tend to be weaker at randomization performance. Organizations deploying AI in climate science (or other high-stakes domains) need to explicitly decide which explanation properties matter most for their use case rather than assuming all desirable qualities can be achieved together.

Architecture choices affect explanation quality: The paper demonstrates that the same XAI method performs differently depending on whether a simpler neural network (MLP) or more complex one (CNN) is used, particularly for robustness, complexity, and localization. This highlights that responsible AI deployment requires task-specific evaluation rather than generic best practices.