Applications of machine learning methods for photonics and non-Hermitian physics

• Main Author: Zhu, Changyan
• Other Authors: Chong Yidong
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Physics; Photonics; Non-Hermitian; Machine learning
• Online Access: https://hdl.handle.net/10356/173955
• Institution: Nanyang Technological University

The recent advances in machine learning and related techniques have arisen their application in different areas. In physics, especially in photonics, Machine learning learns from the dataset and provide an accurate description of mapping between different physical variables. Therefore, they are quite powerful in physics research. This thesis explores various machine learning algorithms for photonics and non-Hermitian physics. Chapter 1 introduces the interplay between machine learning and physics. In particular, dense neural networks and convolutional neural networks are explained in detail. Chapter 2 compares the efficiency of dense neural network (DNN), U-Net, and VGG-net for multi-mode fiber (MMF) image reconstruction, with DNN emerging as the most suitable due to its ability to consider non-local features. Chapter 3 introduces an intelligent real-time and self-adaptive terahertz beamforming scheme based on neural networks, demonstrating accurate beam steering and high generalizability. Chapter 4 uses a neural network with regularization for the reconstruction of random spectrometer signals, outperforming traditional matrix inversion methods in terms of bandwidth and accuracy. A compact, tunable spectrometer is also developed. Chapter 5 extends the exploration to non-Hermitian physics, identifying topological variants and Non-Hermitian Skin Effects (NHSE) using unsupervised learning methods. The potential applications of NHSE in quantum amplifiers and the continuum of bound states in non-Hermitian lattices are also discussed. Chapter 6 summarizes the entire thesis and discusses potential directions for future research based on the current thesis.