Machine-learning based semiconductor device characterisation and behavioural modelling

Yuan, Weihao ORCID: https://orcid.org/0000-0001-9857-574X 2026. Machine-learning based semiconductor device characterisation and behavioural modelling. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Download (16MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (218kB)

Abstract

This thesis investigates methods to improve the efficiency of behavioural modelling workflows for Radio Frequency (RF) power amplifiers (PAs) using load-pull measure ments. In practical design processes, both large-scale simulations and measurements can consume significant time, making model development and design optimisation slower. To improve simulation efficiency, an Artificial Neural Network (ANN) model im plementation framework is developed in a computer-aided design environment based on an existing Cardiff Model (CM) implementation. The framework introduces a dynamic function-generation framework within the ADS schematic, allowing the ANN model implementation to be adapted according to the required model struc ture. Based on this framework, a systematic runtime comparison between ANN model and CM implementations is presented under load-pull conditions of varying complexity. The results show that schematic structure has a strong effect on sim ulation cost. For the fundamental-frequency simulation considered in this study, removing calculations that were redundant for the selected model configuration re duced the runtime of the original CM implementation by approximately 99% while maintaining modelling accuracy. In addition, this thesis investigates systematic (Type B) uncertainties in load-pull measurement systems that may remain hidden after calibration. A novel method is developed to characterise and assess these uncertainties. A measurement system integrity assessment workflow is introduced, enabling early detection of abnormal deviations before they affect behavioural model extraction and device performance evaluation. The results show that the measurement-domain threshold is exceeded by system atic distortions of around 1-2% of the incident-wave magnitude, while the model domain threshold is exceeded at around 2.6%. These results indicate that proper model implementation can significantly reduce simulation time, while early detec tion of systematic distortions can help reduce the time wasting on re-measurement and recalibration. Overall, this work provides a new perspective on improving both computational and experimental efficiency in load-pull based behavioural modelling, contributing to faster and more reliable RF PA design workflows.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Engineering
Uncontrolled Keywords:	1. RF Power Amplifier 2. Behavioural Modelling 3. Artificial Neural Network 4. Cardiff Model 5. Load-Pull Measurement 6. Measurement Uncertainty
Date of First Compliant Deposit:	29 September 2026
Last Modified:	29 Sep 2026 15:27
URI:	https://orca.cardiff.ac.uk/id/eprint/189873

Actions (repository staff only)

Edit Item