Towards machine learning-assisted electronic design automation: microwave filter, power amplifier, and semiconductor device

Xue, Liyuan (2024) Towards machine learning-assisted electronic design automation: microwave filter, power amplifier, and semiconductor device. PhD thesis, University of Glasgow.

Abstract

Over the decades of development, electronic design automation (EDA) has been widely applied in most electronic design problems, especially in advanced and sophisticated digital systems. In contrast, the degree of automation for distributed-element circuits, e.g. microwave or millimeter-wave (mm-wave) devices characterized by electromagnetic (EM) simulations, and semiconductor devices characterized by technology computeraided design (TCAD) simulations, is still very limited. Two challenges are especially notable. First, both EM and TCAD simulations are computationally expensive. Second, some design problems in these fields are highly parameter-sensitive with many local optimal solutions. Consequently, fully algorithmic EDA in these fields is still in its infancy, especially incorporating with advances of machine learning (ML) or artificial intelligence (AI) techniques for higher automation levels.

The objective of this thesis is accordingly to develop a more generic and effective framework (than hitherto) for design automation in these fields, assisted by cutting-edge progress in ML. Three representative circuits/devices are selected for investigation: microwave filter, monolithic microwave integrated circuit (MMIC) power amplifier (PA), and semiconductor devices. Beginning with a brief introduction of EDA, basic concepts of relevant optimization algorithms and ML techniques are brought in subsequently, then each topic is unfolded by a comprehensive literature review followed by details of the proposed methodology, experimental results, and comparisons. Specifically,

• Microwave Filter: A design automation method composed of two-phase design optimization is proposed for three-dimensional microwave filters. In each phase, the bespoke objective functions and optimization algorithm are proposed to improve the robustness and success rate. By incorporating with a programmable initial design synthesis, the proposed methodology enables the first unsupervised design automation without human intervention.

• MMIC PA: An efficient layout-level automated design methodology is proposed for MMIC PAs, supporting holistic characterization with EM, small- and largesignal simulations and being compatible with most foundry process design kits. Bayesian neural networks are integrated with novel hybrid local and global search strategies. Two MMIC PAs—a balanced Class-AB PA and a wideband Doherty PA—were successfully synthesized with the later taped out for manufacturing.

• Semiconductor Device: An attempt towards algorithmic design optimization for semiconductor devices is presented through two case studies. The first optimized the epitaxial layer of a commercial III-V pHEMT for higher cut-off and maximum oscillation frequency over terahertz, achieving a 30% and 57% improvement, respectively. The second study proposed the concept of device circuit co-optimization, enhancing the performance of a planar CMOS-based inverter to outperform several reported devices with advanced technologies.

In conclusion, this thesis investigated ML-assisted EDA within the three aforementioned areas. The research outcomes demonstrate significant improvements in design efficiency, performance, and versatility. This work paves the way for further research into higher degrees of design automation, facilitating the emergence of the upcoming AI-driven EDA era.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Supervisor's Name: Liu, Professor Bo, Onireti, Dr. Oluwakayode and Imran, Professor Muhammad
• Date of Award: 2024
• Depositing User: Theses Team
• Unique ID: glathesis:2024-84810
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 14 Jan 2025 15:08
• Last Modified: 14 Jan 2025 15:08
• Thesis DOI: 10.5525/gla.thesis.84810
• URI: https://theses.gla.ac.uk/id/eprint/84810

Related URLs

• Article DOI
• Enlighten Publications Record
• Article DOI
• Article DOI
• Article DOI