Machine learning-aided design of additively manufactured ridge gap waveguide components

Abstract

The rapid evolution of next-generation wireless systems poses significant challenges for the design of compact, high-performance passive components and massive-MIMO antenna arrays. Conventional workflows for ridge gap waveguide (RGW) and printed ridge gap waveguide (PRGW) unit cells rely on iterative full-wave EM simulations and parametric studies to locate stop-band frequencies, resulting in prohibitive computational cost, long turn-around times, and limited generalizability of analytical dispersion formulas. Meanwhile, advances in computational electromagnetics (CEM) including MoM, FEM, and FDTD solvers have eased model fidelity but not the excessive memory and processing demands required to refine designs by hand. Motivated by the success of machine learning (ML) in other fields and the growing volume of AI-based EM research, this thesis develops a fully automated, fabrication-aware synthesis framework for RGW/PRGW unit cells that leverages supervised learning to replace manual trial-and-error. A high-fidelity dataset covering 3–300 GHz and accounting for geometric variations related to additive manufacturing and dielectric materials was generated via automated EM simulations. Genetic programming (GP) and artificial neural networks (ANN) were then implemented, benchmarked, and optimized to learn both forward (geometry → performance) and inverse (performance → geometry) mappings with near-full-wave accuracy and drastically reduced run-times (minutes vs. hours/days). Key contributions include: - The first end-to-end ML-driven synthesis framework for metallic RGW/PRGW unit cells. - A scalable, fabrication-aware EM dataset enabling data-driven modeling across 3–300 GHz. - Systematic comparison of GP and ANN inverse-design approaches using MSE, MAE, MAPE, and timing metrics. - Design, 3D-printing, and experimental validation of wideband, highly isolated, pattern-reconfigurable PRGW-based MIMO antennas, including an origami-inspired 8 × 8 array with multi-permittivity dielectric resonators. The results demonstrate that ML-aided synthesis can overcome brute-force EM iteration, offering a scalable path toward compact, cost-effective System-on-Package modules for future mm-wave communications. The thesis concludes with a discussion of limitations and outlines promising directions for reinforcement learning, advanced data generation, and inverse modeling in EM structure design.

Date	8 Sept 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Ghyslain Gagnon (Supervisor) & Dominic Deslandes (Co-supervisor)