Quantitative ultrasonic inversion of coating bond strength based on a physics-informed neural network

Abstract

The nondestructive quantitative evaluation of the bond strength of metal coatings on marine equipment is crucial for ensuring their safe service. However, due to the inherent high acoustic attenuation of metal coatings and the limited number of destructive pull-off test samples, traditional ultrasonic techniques and purely data-driven deep learning models find it difficult to balance physical interpretability and high-fidelity prediction. To address this challenge, this paper proposes a novel quantitative characterization method that integrates wideband pulse compression surface waves with physics-informed neural network (PINN). This study first reveals the acoustic response mechanism underlying the interfacial degradation between the coating and the substrate. It clarifies that interfacial microvoids and air gaps induce geometric dispersion and energy dissipation hysteresis in surface waves. Based on these insights, the time-domain main lobe width and the signal residual vibration decay coefficient are extracted as key indicators for characterizing interfacial degradation. Subsequently, a dual-driven PINN architecture was constructed by combining the data loss with a physics-based regularization term formulated from a simplified interfacial acoustic relationship. In this model, C1 and C2 are defined as trainable effective model coefficients and are jointly optimized with the neural-network weights. Experimental results indicate that the proposed model achieves quantitative inversion of the interfacial bond strength for complex metal coatings, yielding a fitting coefficient R2 of 0.98 and a root mean square error (RMSE) of 0.51 MPa. This study establishes a quantitative relationship between interfacial acoustic degradation features and macroscopic mechanical strength, providing a physics-informed approach for the nondestructive evaluation of the service quality of metal coatings on marine equipment.