Defect estimation using surrogate-based Monte Carlo Bayesian optimization

This study presents a novel composite AI approach that integrates Monte Carlo Bayesian optimization with a fast neural network surrogate of acoustic wave propagation to estimate the depth and width of defects in noisy nondestructive testing (NDT). The surrogate model, trained on finite element simulations, predicts von Mises stress from inputs of excitation width, material coordinates, and propagation time—delivering outputs more than 70,000 times faster than direct numerical analysis. By systematically adjusting the surrogate inputs to match arbitrarily given (target) outputs, the optimizer identifies the depth and width of the defect without relying on signal-processing algorithms. Experimental evaluations demonstrate over 98.5% and 90% accuracy in estimating defect depth and width (with less than 1 mm error, respectively), even in challenging low signal-to-noise conditions. Comparative evaluations using different similarity metrics ( R 2 , r , cosine similarity, inverted MSE, or inverted MAE) indicate that Monte Carlo Bayesian optimization guided by the R 2 -metric consistently delivers robust and stable performance, underscoring its resilience to high noise levels. These results support the potential for real-time, high-accuracy ultrasonic testing for defects while simultaneously reducing computational demands and minimizing the need for specialized, domain-specific algorithm development. • Combines Monte Carlo Bayesian optimization with a neural network surrogate model. • Surrogate model over 70,000 times faster than finite element simulations. • Achieves > 98.5% depth and 90% width estimation accuracy within 1 mm error. • Method robust under poor SNR conditions; effective in noisy environments. • Similarity metrics like R 2 perform strongly across various noise levels.