Bayesian neural networks-based surrogate model-assisted evolutionary algorithms and their applications to microwave antenna design

Liu, Yushi (2024) Bayesian neural networks-based surrogate model-assisted evolutionary algorithms and their applications to microwave antenna design. PhD thesis, University of Glasgow.

Abstract

Gaussian process (GP) is a widely used machine learning model for optimising online surrogate model-assisted antenna design. Despite many successes, two improvements are essential for the GP-based antenna global optimisation methods. First, the GP model training costs when there are many design variables and specifications. Second is the convergence speed (i.e., the number of necessary electromagnetic (EM) simulations to obtain high-performance designs). In both aspects, the state-of-the-art GPbased methods show practical but undesirable performance, particularly for optimising modern antennas, which often have many design variables and specifications.

Therefore, a behavioural study of a potential surrogate model alternative, Bayesian neural network (BNN), which has yet to be paid attention to, is presented in this thesis. Through empirical studies, the properties of the BNNs, their co-work with prescreening methods, and their comparison with other machine learning model alternatives are investigated with a typical surrogate model-assisted evolutionary algorithm (SAEA) model management framework. The behaviour of BNNs regarding surrogate model prediction accuracy, the availability of prediction uncertainty estimation, and the training cost are demonstrated in the experiments, showing the potential of BNNs to be a competitive alternative for online surrogate model-assisted antenna design optimisation.

Thus, this thesis presents an upgraded antenna design optimisation method called selfadaptive Bayesian neural networks surrogate model-assisted differential evolution for antenna design exploration (SB-SADEA). The key innovations include (1) the introduction of the BNNs-based antenna surrogate modelling method into the research area, replacing widely used GP modelling, and (2) a bespoke self-adaptive lower confidence bound (LCB) method for antenna design landscape making use of the BNNs-based antenna surrogate model. A slotted monopole antenna (ultra-wideband, 40% area reduced), a 5G mm-wave antenna (20 design variables, 12 design specifications, four operating bands), a sub-6 GHz outdoor base station antenna (23 design variables, 18 design specifications including S-parameters, front-to-back ratio and half-power beam-width) and a microstrip patch antenna (quasi-digitally coded, 62 design variables) are used to test the performance of SB-SADEA in the thesis. The results show considerable improvement in convergence speed and machine learning cost compared with the state-of-the-art GP-based antenna global optimisation methods.

Furthermore, the proposed BNN-based SAEA has been tested against global optimisation applications on a broader scope. A supercontinuum generation waveguide and a holistic radar signal processing and classification system (12 design variables, including data pre-processing and feature extraction parameters in binary, continuous and discrete forms) are used to test the performance of the proposed algorithm. The experiments show that the proposed algorithm is efficient in optimising not only antenna structures but also components, structures and systems in other domains. Moreover, the proposed algorithm is compatible with discrete and categorical design variables.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Additional Information: Supported by funding from MathWorks Inc. and the University of Glasgow.
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science; T Technology > T Technology (General)
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Funder's Name: MathWorks Inc.
• Supervisor's Name: Liu, Professor Bo, Ur-Rehman, Dr. Masood and Imran, Muhammad
• Date of Award: 2024
• Depositing User: Theses Team
• Unique ID: glathesis:2024-84608
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 09 Oct 2024 10:47
• Last Modified: 10 Oct 2024 08:24
• Thesis DOI: 10.5525/gla.thesis.84608
• URI: https://theses.gla.ac.uk/id/eprint/84608

Related URLs

• Article DOI
• Article DOI
• Article DOI