Felicia Guo

EECS Department, University of California, Berkeley

Technical Report No. UCB/

December 1, 2025

Analog and mixed-signal (AMS) integrated circuits (ICs) lie at the core of modern computing and communications systems. However, despite the continued rise in design complexity, advances in AMS automation remain limited. This dissertation presents a series of methods aimed at advancing automation in AMS circuit design.

The first part of this work examines codified, generator-based design methodologies through the development of an analog-to-digital converter (ADC) design flow and generator, implemented in the Berkeley Analog Generator (BAG3++) framework, that produces ADCs with resolutions from 4 to 8 bits. As part of this effort, we developed the SkyWater 130nm PDK plugin to enable BAG3++ for open-source AMS design. While this approach enables reproducible, parameterized design flows, its reliance on fixed topologies highlights a limitation of template-based design. To address this challenge, we formulate analog circuit synthesis as a graph-generation problem and introduce a reinforcement-learning (RL) framework capable of autonomously constructing circuit topologies guided by simulation-based feedback. This framework demonstrates an expert-aligned method for generalized circuit generation and is validated across multiple AMS design tasks spanning both linear and nonlinear designs, including a ring oscillator, comparator, and operational transconductance amplifier. Across these tasks, 100% of generated netlists are structurally valid by construction, and over 90% of generated circuits exhibit the intended functionality. Finally, we explore extensions of these learning-based methods toward broader automation, demonstrating 1) the sizing and stability compensation of a two-stage operational amplifier ultimately achieving high gain with greater than 60° phase margin and 2) the potential for interfacing large language models (LLMs) with codified generator application process interfaces (APIs) for layout generation. Together, these contributions represent a step toward adaptive, performance-driven AMS design automation.

Advisors: Borivoje Nikolic

---

BibTeX citation:

@phdthesis{Guo:32006,
    Author= {Guo, Felicia},
    Title= {Towards Automated Analog-Mixed Signal Circuit Design: Reinforcement Learning and Codified Generator Frameworks},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {Dec},
    Number= {UCB/},
    Abstract= {Analog and mixed-signal (AMS) integrated circuits (ICs) lie at the core of modern computing and communications systems. However, despite the continued rise in design complexity, advances in AMS automation remain limited. This dissertation presents a series of methods aimed at advancing automation in AMS circuit design.

The first part of this work examines codified, generator-based design methodologies through the development of an analog-to-digital converter (ADC) design flow and generator, implemented in the Berkeley Analog Generator (BAG3++) framework, that produces ADCs with resolutions from 4 to 8 bits. As part of this effort, we developed the SkyWater 130nm PDK plugin to enable BAG3++ for open-source AMS design. While this approach enables reproducible, parameterized design flows, its reliance on fixed topologies highlights a limitation of template-based design. To address this challenge, we formulate analog circuit synthesis as a graph-generation problem and introduce a reinforcement-learning (RL) framework capable of autonomously constructing circuit topologies guided by simulation-based feedback. This framework demonstrates an expert-aligned method for generalized circuit generation and is validated across multiple AMS design tasks spanning both linear and nonlinear designs, including a ring oscillator, comparator, and operational transconductance amplifier. Across these tasks, 100% of generated netlists are structurally valid by construction, and over 90% of generated circuits exhibit the intended functionality. Finally, we explore extensions of these learning-based methods toward broader automation, demonstrating 1) the sizing and stability compensation of a two-stage operational amplifier ultimately achieving high gain with greater than 60° phase margin and 2) the potential for interfacing large language models (LLMs) with codified generator application process interfaces (APIs) for layout generation. Together, these contributions represent a step toward adaptive, performance-driven AMS design automation.},
}

EndNote citation:

%0 Thesis
%A Guo, Felicia 
%T Towards Automated Analog-Mixed Signal Circuit Design: Reinforcement Learning and Codified Generator Frameworks
%I EECS Department, University of California, Berkeley
%D 2025
%8 December 1
%@ UCB/
%F Guo:32006