Physics-Constrained Transfer Learning for Self-Assessment of Terahertz Thickness Measurement of Thermal Barrier Coatings

Terahertz (THz) thickness measurements of thermal barrier coatings vary within a certain range in terms of accuracy and occasionally generate unsatisfying results. To identify such results, a physics-constrained transfer learning framework is presented to achieve self-assessment of measurement results. First, an advanced neural network is employed to estimate thickness, roughness, and refractive index, followed by inputting them into the analytical model to generate simulated traces, which are subtracted by the actual signals to obtain the residual data. Then, we find that the residuals of the first two peaks from experimental and simulated THz signals are similar parts, and they are correlated to the error through the phase factor in THz physics. Thus, a physics-constrained layer is presented to amplify the residuals of the first two peaks to improve the similarity. Finally, the trained physics-constrained transfer learning can achieve satisfactory accuracy for assessing the thickness measurement results of actual specimens.