Yifeng Bie

Topic

Machine Learning for Optical Design and Analysis

Department of Electrical and Computer Engineering

Date & location

• Tuesday, September 15, 2026

• 8:30 A.M.

• Engineering Office Wing

• Room 502 & Virtual Defence

Reviewers

Supervisory Committee

• Dr. Tao Lu, Department of Electrical and Computer Engineering, University of Victoria (Supervisor) 

• Dr. Issa Traoré, Department of Electrical and Computer Engineering, UVic (Member)

• Dr. Irina Paci, Department of Chemistry, UVic (Outside Member) 

External Examiner

• Dr. Tian Yang, Department of Electrical Engineering, Shanghai Jiaotong University 

Chair of Oral Examination

• Dr. Sandra Gibbons, School of Exercise, Science, Physical and Health Education, UVic

   

Abstract

Optical spectroscopy and photonic-device modelling rely on computational methods to interpret complex data. In quantitative optical absorption spectroscopy, for example, measured spectra are related to the composition and concentrations of chemical mixtures. In photonic-device modal analysis, numerical eigensolvers produce many electromagnetic modes whose physical identities must be determined as the device geometry changes. This dissertation presents machine-learning methods for these two distinct problems: data-efficient quantification of multi-component absorption spectra and mode identification and tracking in coupled waveguide-disk cavities. 

The first project investigates dimensionality reduction and concentration prediction for multi-component absorption spectra. Principal component analysis is used to study the relationship between spectral dimensionality, the number of absorbing species, and measurement quality. Functional principal component analysis is also introduced to represent absorption spectra as continuous functions of wavelength and to preserve their characteristic spectral structure. The resulting functional components and scores are examined in relation to molecular absorption features and species concentrations, providing a physical basis for quantitative prediction. Building on these relationships, several functional-component-based models are developed, including a nearly training-free approach designed to reduce dependence on large calibration datasets. The proposed methods are evaluated using both simulated multi-gas spectra and experimental dye-mixture spectra. The results demonstrate that physically meaningful functional representations can reduce data dimensionality while maintaining reliable concentration prediction, including in conditions beyond the range represented by the calibration data. 

The second project develops a compact framework for tracking optical modes in a coupled waveguide-cavity system as the geometry is varied. To track mode family upon geometry sweeping, each mode is represented using a small set of descriptors derived from its eigenfrequency and electromagnetic field distribution, including resonance behaviour, quality factor, energy confinement, spatial localization, and polarization. A hierarchical unsupervised procedure is developed to assign consistent mode-family labels across changing geometries. The method provides a reliable and interpretable approach to photonic mode tracking, which paves the way for efficient mode matching modelling of photonic devices.