Nondestructive measurement of elastic stiffness in anisotropic aluminium alloys using resonance frequency-guided machine learning models

This study introduces a novel, non-destructive methodology for determining the elastic stiffness of orthotropic aluminum sheets with unknown material axis orientation. The approach integrates resonance frequency spectra with an inverse numerical model and machine learning (ML) techniques – Random Forest (RF), Shallow Neural Networks (SNN), and Extreme Gradient Boosting (XGB). Among the models, XGB delivered the most accurate results, with R2 values of 0.947 for E₁₁, 0.938 for E₂₂, 0.930 for G₁₂, 0.912 for Poisson’s ratio (ν₁₂), and 0.904 for the principal material axis angle (θ). RF followed closely, achieving R2 values of 0.924 (E₁₁), 0.905 (E₂₂), 0.909 (G₁₂), 0.895 (ν₁₂), and 0.879 (θ). SNN showed slightly lower performance, with R2 values of 0.901, 0.871, 0.866, 0.849, and 0.839 for the same properties, respectively. These results indicate that all models were more effective in predicting elastic moduli (E₁₁, E₂₂, G₁₂) than in estimating Poisson’s ratio and material orientation. Additionally, increasing resonance spectra from multiple sampling points improved model accuracy initially, but beyond a threshold, redundant overlapping spectral data diminished the overall improvement. Finally, a case study demonstrates how different anisotropic properties – arising from variations in rolling conditions – impact the resonance spectrum, validating the method’s sensitivity to directional material properties.