Dual-Stage Fourier Neural Operator: Mitigating Spectral Bias for Enhanced Ultrasonic Guided Wave Corrosion Imaging

Deep neural networks exhibit spectral bias in data-driven ultrasonic guided wave corrosion imaging, preferentially learning low-frequency components while struggling with high-frequency features critical for defect boundary delineation. This paper proposes a Dual-Stage Fourier Neural Operator (DS-FNO) framework designed to mitigate this limitation. The first stage recovers dominant low-frequency velocity components, while the second stage employs residual learning to capture fine-scale spectral content. Numerical experiments on aluminum plate structures validate the feasibility and effectiveness of the proposed dual-stage architecture for enhanced corrosion imaging.