A method for improving ray distortion in ultrasonic weld simulation based on a ray-guided Fourier neural operator

Abstract

Forward simulation of ultrasonic nondestructive testing in austenitic stainless steel welds remains challenging due to the material’s severe anisotropy and heterogeneity. Traditional ray-tracing methods are highly efficient but suffer from amplitude singularities in focal regions, whereas full-wavefield finite element methods are computationally prohibitive for large-scale parametric analysis. To break this efficiency-accuracy trade-off, this study proposes a ray-guided Fourier neural operator (FNO) framework. Rather than learning wave physics from scratch, the proposed framework employs high-frequency asymptotic ray solutions as kinematic physical priors. Through a dual-channel input and residual learning architecture, the FNO focuses exclusively on correcting the local dynamic distortions of the ray model in focal zones and high-gradient regions. Trained on a parameterized weld dataset, the model incorporates a Sobolev space-based gradient constraint loss to ensure physical continuity and generalization to unseen weld geometries. Numerical validations demonstrate that the proposed method significantly mitigates prediction errors in focal regions of asymmetric welds—reducing errors from 34.6% to 7.2%—while maintaining sub-second inference speeds comparable to standard ray tracing. This framework successfully bridges the gap between fast asymptotic analysis and high-fidelity full-wave calculations, enabling efficient large-scale Monte Carlo uncertainty quantification for nuclear power structures.