Neural Network Attribution Methods for Problems in Geoscience: A Novel Synthetic Benchmark Dataset

XAI methods lack objective evaluation standards: The paper identifies a critical gap in responsible AI—most benchmark datasets used to test explanation methods like MNIST and ImageNet lack "objective, theoretically derived ground truth for the attribution," making it difficult to objectively assess whether these explanation methods are actually reliable.

Neural networks in geoscience need better interpretability: While neural networks are increasingly used for geoscience problems, their "complex and nonlinear structure makes the interpretation of their predictions difficult," which directly undermines scientific trust and prevents researchers from gaining physical insights into what the models are actually learning.

A new solution: synthetic benchmarks with known ground truth: The researchers developed a framework using "additively separable functions" to create benchmark datasets where the correct attribution is known in advance, allowing for objective evaluation of different explanation methods rather than subjective assessment.

Different XAI methods have inconsistent performance: By testing multiple explanation methods against their benchmark dataset, the authors identified "examples where specific XAI methods perform well or poorly," suggesting that no single explanation method is universally reliable—a crucial finding for practitioners choosing which method to trust.

Geoscience-specific benchmarks are needed: The paper addresses a practical gap by creating benchmarks "specifically designed for problems in geosciences," recognizing that general-purpose benchmarks may not adequately test explanation methods for domain-specific applications.