Session: 21-01: Poster Session

Paper Number: 98008

98008 - Machine Learning Inversion to Experimental Dispersion Curves for Characterizing Thin Coatings 

This research uses guided Lamb waves to characterize the thickness and uniformity of thin coatings, specifically when the coating material properties are similar to those of their base plate. The nondestructive characterization of such coatings is broadly applicable to industries which require specifically tailored surface properties and precision identification of defects, such as the aerospace and nuclear energy fields. Methods exist to invert dispersion curves to coating properties, but such methods require fitting to theoretical curves [1] or the utilization of the Global Matrix Method[2]. As such, this research proposes the utilization of a machine learning inversion scheme using training data obtained from Abaqus FEM simulation. This method offers usage of experimental curves as a stand-in for analytical models, enabling efficient inversion to coating thickness.

The first task in this research is to experimentally obtain dispersion curves for thinly coated samples. A stainless-steel sample with a 20-micron-thick zinc coating is to be used in this experiment, as well as a chromium-coated zircalloy-4 sample. Analytical and finite element simulated dispersion curves suggest that high frequency excitation maximizes the separation of low-order modes for the plate and coating. Experimentally obtaining dispersion curves at such high frequencies requires novel wedge design and maintenance of strict experimental conditions. The second task of this research is to perform an inversion of said dispersion curves. Non-Maximum Suppression (NMS) was used on dispersion curves generated from FEM simulations where coating thickness was varied between 10 µm and 600µm. The data selected from NMS undergoes a basic curve-fitting algorithm, and the coefficients of the resultant fit are used as training data for a machine learning model. NMS will then be performed on experimental dispersion curves and the result will be subject to the tuned machine learning classifier, yielding a characterization of whether the thickness of the coating is within an acceptable tolerance. Training on the Cr-Zr4 model with simulated data suggests a prediction accuracy in excess of 90%.

References

[1] Jens Christian Stolzenburg, Jacek Jarzynski, and Laurence J. Jacobs , "Near field inversion method to determine the material properties of a layered media", AIP Conference Proceedings 615, 1415-1422 (2002) https://doi.org/10.1063/1.1472960

[2] Koreck, J., Valle, C., Qu, J. et al. Computational Characterization of Adhesive Layer Properties Using Guided Waves in Bonded Plates. J Nondestruct Eval 26, 97–105 (2007). https://doi.org/10.1007/s10921-007-0024-y

Presenting Author: Charles Tenorio Georgia Institute of Technology