Noise robust real-time focus detection with deep learning for ultra-fast laser micromachining

Abstract

In this thesis, different types of machine learning models are provided for ultra-fast laser micromachining system to actively control the focusing of light on the processing material by detecting the reflected light. These different types of models are tested for both experimental and theoretical approaches. For the experimental approach, four different machine learning models are explored. These models were tested for mirror, silicon, steel, and copper samples. The proposed machine learning models offer real-time control with over 90\% accuracy. For the simulation, noise at the material surface and the detection system are considered. The noise simulation, including the laser micromachining system, is done using Fourier optics and signal processing. Noise levels at the material surface are determined by laser scanning microscope measurements of experimental samples, and the commercial detection camera noises are considered for the detection noise. Convolutional neural network models are used for focus control in the simulation. Depending on the noise level, the proposed model achieves above 95\% accuracy.