OperatorSHAP Provides Fast, Accurate Shapley Values for Neural Operators.

Understanding the predictions of machine learning models is paramount in safety-critical physical applications, such as structural load assessment or weather forecasting. Shapley values are a highly desirable attribution method due to their robust theoretical properties, but their high computational cost during inference has limited their practical deployment. Existing amortized explainers, like FastSHAP, are often restricted to homogeneous inputs, which is problematic for physical data that frequently originates from irregular grids and geometries. Researchers have introduced OperatorSHAP, a groundbreaking grid-agnostic attribution method specifically designed for neural operators. This method includes a unique training procedure that allows for FastSHAP-like explainers to be applied to these complex models. OperatorSHAP establishes a comprehensive theoretical framework for attributions within function space, directly connecting to Aumann-Shapley values, which are essential for continuous input spaces. A key advantage of OperatorSHAP is its ability to generate explanations that are consistent across different resolutions and can transfer across varying grid sizes without requiring the model to be retrained. This significantly enhances its utility for real-world physical applications, where data resolution and structure can vary. The method promises to make model interpretability more accessible and efficient for critical scientific and engineering domains.