Fourier Neural Operators for Reissner–Mindlin Plate Mechanics: Shear Locking, Mixed Formulations, and Benchmarking Protocols

This paper analyzes the use of Fourier Neural Operators as surrogate solvers for Reissner–Mindlin plate bending with a specific focus on the thin-plate regime, where shear locking dominates both numerical discretizations and learning-based models. It argues that neural operators trained on locking-contaminated finite element data inherit artificial stiffness and therefore cannot be physically reliable. To address this, the paper synthesizes locking-free data generation strategies based on mixed finite element formulations, reduced shear treatments, and validation against analytical benchmarks. It further examines how spectral bias in Fourier-based neural operators interacts with high-frequency shear boundary layers and outlines mitigation strategies such as Fourier-feature lifting, hard boundary enforcement via distance functions, and mixed-variable loss formulations. Finally, it proposes benchmarking protocols that explicitly test thickness-dependent stability, accuracy in both displacement and derived quantities, and computational speedup, positioning FNOs as viable real-time surrogates only when grounded in locking-free variational structure.