LSWNet: A physics-informed neural network for ultrasonic wavefield prediction and elastic constant inversion in...

Accurate determination of elastic constants is crucial for reliable ultrasonic defect detection in carbon fiber reinforced plastic (CFRP). However, non-destructive in-situ characterization of these constants, particularly via full-waveform inversion techniques, is hindered by the computational cost of wavefield simulations. Based on physics-informed neural networks (PINNs), a novel longitudinal and shear wavefield net (LSWNet) method is proposed for the forward wavefield prediction and inversion of ultrasonic waves in a unidirectional CFRP. The longitudinal and shear wave component fields at two moments, ultrasonic measurement data, and the 2D elastic wave equations of isotropic and anisotropic planes for unidirectional CFRP are embedded as physical constraint conditions to predict wavefields and elastic constants. For the inversion of elastic constants, ultrasonic data recorded by a linear phased array on the CFRP surface serve as input, while the LSWNet outputs C₆₆, C₁₃ and C₄₄. To accelerate convergence in large-scale models, weights and biases learned from training on small-scale structures are transferred. The proposed method has been verified through both finite element simulation and experiments. The mean squared errors between the predicted wavefields by PINNs and those obtained from finite element simulation do not exceed 3.2 × 10^-3, and the obtained elastic constants are close to the actual values. Furthermore, the elastic constants obtained via LSWNet are successfully applied to total focusing method, thereby enabling high-resolution detection of delamination damage. Consequently, the proposed method is capable of resolving forward and inverse issues associated with unidirectional CFRP ultrasonic wavefields, as well as in-situ characterization of elastic constants and damage imaging.