Deep-learning approach for predicting crystalline phase distribution of femtosecond laser-processed silicon

Abstract

In this study, two deep-learning models are presented to predict the crystalline phases of femtosecond laser-processed silicon. To obtain the datasets, single-crystal silicon was processed by a femtosecond laser using 49 different combinations of laser fluence and scanning speed, and for each specimen, Raman spectra were measured at 22,500 locations inside a square domain. The first model was trained to classify the Raman spectra of silicon into six silicon phases. By applying the model to the entire surface of a silicon specimen in batches, the silicon phase distribution was visualized in RGB color values, with each color representing a particular silicon phase. Using the classification results of the 49 specimens obtained by the first model, the second model was developed to predict the silicon phase distribution image from the inputs of the laser fluence and scanning speed. The average prediction accuracy of the second model was 86.60%.

Keywords

• Femtosecond laser annealing of silicon
• Silicon crystalline phase
• Raman spectra
• Deep learning
• Generative adversarial network