Physics‐Informed Agent‐Analytics Framework With Causal Validation for Crease Prediction in Multilayer Flexible Cover Stacks

ABSTRACT We propose an agent‐analytics framework for data‐driven crease mitigation in multilayer flexible cover stacks, where an LLM‐assisted workflow supports a human domain expert in formalizing engineering hypotheses into mechanics‐grounded descriptors and in organizing ensemble modeling, explainability, and adjustment‐based causal‐consistency analysis pipelines. Starting from neutral‐axis mechanics and the parallel axis theorem, the framework systematically generates 14 physics‐informed descriptors organized into five mechanism groups: geometry, stiffness balance, elastic energy, viscoelasticity, and interface mechanics. Through a three‐stage evidence hierarchy—ensemble SHAP importance, bootstrap pathway analysis under progressive adjustment, and progressive confounder adjustment—we identify three independently robust descriptors while demonstrating that the remaining descriptors operate through these base mechanisms. Evaluated on 63 physically fabricated stack variants, the 14‐model ensemble achieves a best cross‐validated R 2 = 0.71. The progression from ratio‐based stiffness symmetry (Tier 3) to absolute moment imbalance (Tier 1) is validated by both improved prediction and evidence‐tier promotion, confirming a shared stiffness‐balance mechanism. A comparative SHAP waterfall analysis reveals that the best‐performing deployable design achieves 39% crease reduction below the dataset mean through geometric compensation rather than extreme material reduction, providing actionable prescriptive guidance for stack design.