Search

The integration of artificial intelligence into materials science has accelerated discovery through iterative workflows that cycle through data acquisition, model refinement, prediction, explanation, and hypothesis-driven experimentation. While explainable artificial intelligence (XAI) methods enhance trust and scientific insight by elucidating model decisions, these explanations are not static. This manuscript introduces the novel concept of explainability drift: the systematic degradation, inconsistency, or divergence in the fidelity, stability, and relevance of XAI-generated explanations across successive iterations of materials AI workflows. Distinct from prediction-focused concept drift, explainability drift arises from evolving data distributions, model updates, feature space expansions, and domain shifts inherent to materials exploration. Through a purely conceptual failure analysis, we delineate the mechanisms underlying explainability drift, including temporal instability in feature attributions, erosion of surrogate model alignment, and semantic misalignment between explanations and emerging material knowledge. Drawing on recent peer-reviewed advances in XAI applications to property prediction, microstructure analysis, and generative design, we synthesize theoretical foundations to highlight why drift undermines iterative efficacy. The proposed conceptual framework organizes explainability drift into multidimensional layers—attributional, structural, and epistemic—offering a structured lens for analyzing failure modes without empirical validation. This framework emphasizes risks such as misguided hypothesis generation, diminished trust in AI-assisted insights, and inefficient navigation of vast materials design spaces. By conceptualizing explainability drift as an intrinsic challenge, the work advocates for theoretical advancements in sustained explainability to support robust, interpretable AI-driven materials innovation.

Interpretability is widely claimed for graph neural networks (GNNs) in metallurgy and alloy property prediction. Yet the term remains ambiguous: authors’ frequently present attention maps that highlight atoms or bonds, then conclude that the model has uncovered physical mechanisms governing strength, ductility, or creep resistance. This boundary/definitional article demonstrates that such claims conflate three fundamentally different concepts—statistical explainability, mechanistic interpretability, and causal interpretability—none of which is automatically satisfied by attention weights or feature-importance scores. Attention maps reveal which inputs the model attends to, but they do not establish why those inputs matter, nor do they predict the outcome of microstructural interventions. Drawing exclusively on the peer-reviewed literature that applies GNNs to metallurgical systems, the article distinguishes statistical correlations captured by attention mechanisms from the physical mechanisms required for trustworthy alloy design. It proposes an operational, multi-level definition of interpretability tailored to metallurgy, specifying validation criteria for each level. Statistical explainability (Level 1) is shown to be the weakest form, limited to correlational insights; mechanistic interpretability (Level 2) reveals internal model computations; and causal interpretability (Level 3) demands experimental or simulated interventions to confirm counterfactual predictions. Attention maps do not equal physical mechanisms. The framework clarifies the boundary between useful visualizations and scientifically actionable knowledge, offering concrete reporting standards for model developers, materials scientists, and journal editors. Adoption of these distinctions will accelerate the transition from black-box property prediction to models that genuinely advance metallurgical understanding and enable physics-informed alloy design.