The seduction of correlation: on the epistemic limits of AI in materials physics

Abstract The growing use of Artificial Intelligence in materials physics has ignited deep transformations in how knowledge is generated, interpreted, and validated. While data-driven models have demonstrated remarkable predictive capabilities, their increasing centrality raises critical questions about the status of physical understanding when theory is no longer the primary framework of inference. This Perspective explores the epistemic implications of substituting causal, mechanistic explanations with statistical correlations, highlighting the risks of conflating predictive accuracy with scientific insight. Drawing on foundational concepts in the philosophy of science, I argue that the integration of AI into materials theory must be guided not only by efficiency or performance metrics but by a commitment to interpretability, falsifiability, and conceptual coherence. Rather than rejecting AI tools, I advocate for their critical incorporation within physically grounded modeling strategies that preserve the explanatory aims of physics. Only through such a reflective synthesis can the seductive power of correlation be harnessed without compromising the epistemological integrity of the discipline.