Antenna modelling is an important tool for engineers and researchers in the field of telecommunications, as it allows for the design and optimisation of antennas in different scenarios and for a variety of applications. However, conventional methods of antenna modelling can be computationally expensive and time-consuming, which can limit the exploration of design space and lead to the inaccurate or even failed in antenna design and optimisation.
With the rapid development of wireless communication technology, antenna design has attracted extensive attention. As device for transmitting and receiving electromagnetic (EM) signals, antenna has a significant impact on the performance of wireless communication systems. Over the past decade, various new antenna and analysis methods have been proposed. Generally, the modelling and analysis of antenna are carried out in EM simulation software such as Computer Simulation Technology (CST) Microwave Studio, High-Frequency Structure Simulator (HFSS), which can be used to model and simulate various kinds of antennas, and the corresponding performance such as reflection coefficient, gain, radiation pattern and impedance of antenna can be directly obtained through simulation. Unfortunately, modern antenna design is more complicated because of the increasing number of design variables, complex structures, and environmental factors. Parametric sweep is an important function of EM simulation software that allows designers to get the information of an antenna under different conditions, the time cost to run an EM simulation for individual candidate solution varies from seconds to minutes, or even several hours. An antenna with complex structure may require thousands of EM simulation to model, the cost in time and computational resources are impractical and unacceptable for most designers and researchers.
To address these challenges, machine learning (ML) methods have been developed and applied to improve the efficiency and accuracy of antenna modelling. These methods involve using ML algorithms to train models on data, which can then be used to predict the performance of antennas for a given set of design variables. This thesis employs and combines different ML-assisted antenna modelling methods to reduce time, cost, and computational intensity in antenna design and accelerate the design process without compromising accuracy. First, quick estimation can be performed using the linear regression (LR) method based on limited data and computational resources to obtain guidance and check the feasibility of an antenna design. Then one of the ANN-based methods can be selected for antenna modelling and optimisation according to the antenna design complexity. These methods can be combined into a systematic antenna design process for modern antenna design. This set of processes can model and optimise antenna for different applications and scenarios with broad ranges of design variables. Compared to EM simulation-based and conventional ML-based antenna design methods. This process can perform accurate antenna modelling using significantly reduced time and computational resources and eliminate unnecessary costs in optimisation, fabrication and testing.
In the first part, a concrete embedded antenna is proposed to mitigate the space occupation and aesthetic problems of indoor dense small cell deployment. The LR method is employed to fast estimate the relationship between antenna performance (radiation efficiency, gain, and input impedance) and embedding ambient (embedding depth and concrete dielectric constant) since the EM simulation-based antenna modelling is time-consuming. The complex mutual coupling between the antenna and the concrete leads to a limited amount of simulated data, and LR can model and predict the performance parameters of the antenna with limited data and a few computing resources. LR can also use limited resources to evaluate the feasibility of antenna design before implementation and fabrication, which can reduce unnecessary overhead and identify potential issues in the antenna. The findings of this study are beneficial to antenna designers for indoor communication concrete embedding antenna design and deployment, as well as communication-friendly building materials.
In the second part, a heuristic algorithm-enhanced artificial neural network (ANN) is proposed to model concrete embedded antenna. The utilisation of ANN can handle the complex and non-linear relationship between inputs and outputs, and it can also make a prediction on antenna performance when new design points are given. A global optimisation algorithm is used to enhance ANN to eliminate local minima issues, and Bayesian regularisation (BR) is employed to improve the network prediction accuracy at new design points. The network accuracy and efficiency are higher than the conventional back-propagation ANN.
The third part proposes a multi-fidelity neural network for antenna modelling and optimisation. Two sources of simulated data are involved and combined to perform antenna modelling with a large amount of cheap and inaccurate models and a small amount of expensive and accurate models. The correlation between two sources of data can be learned adaptively by decomposing the correlation into linear and non-linear components. The feasibility of the approach is validated by three antenna structures, the results show that this method can make prediction for broad ranges of input parameters with satisfactory accuracy; then the surrogate model is directly applied in the optimisation algorithm framework to replace EM simulation to accelerate antenna optimisation procedure.
