Predicting Angle Beam Ultrasonic Inspection Signals with a Physics-Informed Neural Network

Simulation of angle beam ultrasonic inspection supports procedure qualification, probe design, and defect detectability studies, but conventional solvers require a conforming mesh for each configuration and return the solution only at mesh nodes and discrete time steps. Physics-Informed Neural Networks (PINNs) have been applied to ultrasonic wave propagation, but existing formulations solve the wave equation within the propagation medium alone and impose the excitation as a prescribed surface source. The quantity an inspector actually measures; the voltage returned at the transducer terminal; is, therefore, not predicted. This study presents a PINN that couples elastic wave propagation, electrostatics, and piezoelectricity across the complete inspection assembly, comprising the transducer, matching layer, wedge, damping block, and test specimen, and computes the terminal voltage signal directly from an applied excitation. The coupled residuals, together with interface, impedance, fracture, and circuit boundary conditions spanning boundary segments, are embedded in a single loss function and enforced at collocation points without spatial discretization. The framework is applied to an aluminium specimen containing a zero-thickness fracture and validated against both dG-FEM results and experimental measurements. The predicted arrival time of the defect echo agrees with measurement to within 0.1 µs, and the terminal voltage waveform is reproduced with a relative L2 error of 2%. Once trained, the network evaluates the solution at arbitrary points in space and time at a cost of 0.003 ms per point, following a training cost of 1.8 h, making dense evaluation of a trained configuration approximately 20 times faster than the reference solver. Because the network takes only spatial and temporal coordinates as inputs, a change of defect or probe geometry requires retraining; extension to parametric sweeps would require the geometric parameters to be admitted as additional input dimensions. The defect-related reflected signal is approximately 300 times smaller in amplitude than the direct pulse, yet the network recovers approximately 98% of the measured peak reflected amplitude, against 99% for the reference solver.