Inverse Design in Topology Optimization: A Multi-Objective Evolutionary Approach using NGNets and GPR

This thesis presents a novel multi-objective topology optimisation (TO) framework for electromagnetic devices, designed to overcome the prohibitive computational costs of traditional methods. By combining a Differential Evolution (DE) algorithm with Normalised Gaussian Network (NGNet) parametrization, this approach reduces the dimensionality of the design space and naturally generates smooth, manufacturable topologies without explicit spatial filtering. The framework is applied to minimise the radiation Q-factor of an antenna under strict surface area constraints Sf. To enable rapid generation of missing designs along the resulting Pareto front, an extensive dataset of optimised NGNet weight vectors is constructed and utilised to train a Gaussian Process Regression (GPR) inverse model. The results indicate that while a raw GPR model struggles with strict surface constraints, a hybrid DE-corrected GPR method ensures a 100% constraint satisfaction rate while maintaining a robust predictive accuracy, with a relative error strictly under 7% on the Q-factor.