Hybrid metrology investigation combining Raman and ellipsometry spectroscopy applied to in line GeSbTe crystallization measurements and deep learning approaches for accurate prediction

To monitor the Ge-rich GeSbTe crystallization process, ellipsometry and Raman spectroscopy were correlated by Machine Learning using a Neural Network approach. Ellipsometry was selected for being a fast, non-destructive, and in-line metrology technique and Raman spectroscopy was selected for its crystallization monitoring potential.

An experimental ellipsometry/Raman spectroscopy dataset was acquired in a 25–410 °C range. Assuming that the crystallization process is germanium-driven, the crystallinity rate was extracted from fitting the Raman germanium-related modes. Neural Network hybridization was performed to predict the crystallinity rate (output) from the raw ellipsometry spectra (input).

Models trained with these experimental data show poor performance, especially in the 390–410 °C crystallization range. The lack of data was identified to be the main issue. To generate data, the experimental data were independently modeled using numeric temperature laws. Ellipsometry spectra were then generated and labeled with crystallinity rates at any given temperature. The synthetic datasets were then used as a training dataset, leading to a better prediction of the crystallization value, dividing the models' mean squared error by more than ten times.

Finally, the synthetic data-trained models were compared to the experimental data-trained models on an experimental test set. Synthetic data-trained models showed better performance in the crystallization range than the experimental data-trained models (∼2 times lower mean squared prediction errors). The proof of concept of this study is thus validated and could lead to promising potential results in fast crystallization rate prediction of phase change material simply using optical experimental raw measurements.