Artificial neural network assisted spectroscopic ellipsometry data analysis of hydrogenated amorphous silicon thin films

Abstract

Spectroscopic ellipsometry data analysis is typically done with an iterative least squares regression process. It can be resource intensive to generate structural and optical models with good initial parameters to create a high-quality fit. Artificial neural networks (ANNs) are trained on simulated ellipsometric spectra to generate structural and optical model parameters to reduce the human and computational resources required for ellipsometry data analysis. One series of ANNs is trained to generate a structural model, while another series is trained to generate optical and structural parameter values. Two hydrogenated amorphous silicon films (a-Si:H) are deposited on native oxide coated silicon (c-Si) wafers, and one a-Si:H film is deposited on soda-lime glass (SLG). An 81-point map of ellipsometric spectra is measured from each sample and analyzed with traditional least squares regression and the ANN assisted method to validate the approach. The traditional least squares regression process determines that the a-Si:H film on SLG requires an optically distinct surface layer in its structural model that is not needed for the a-Si:H films on c-Si. The ANN assisted method produced the same structural model by determining the substrate and the presence of the surface layer for the experimental data. The structural and optical parameters determined from the ANN assisted method have good agreement with those determined from traditional analysis, and these parameters can be input into the least squares regression to produce the same quality of fit as the traditional analysis generally with less computational time and without human defined initial parameters.