Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

VLM-CAD: VLM-Optimized Collaborative Agent Design Workflow for Analog Circuit Sizing

Guanyuan Pan, Yugui Lin, Tiansheng Zhou, Pietro Liò, Shuai Wang, Yaqi Wang

TL;DR

VLM-CAD addresses the challenge of automatic analog circuit sizing by integrating a Vision-Language Model–driven collaborative agent workflow with schematic interpretation. It combines Image2Net for precise schematic annotation, a multi-agent circuit analysis pipeline, DC bias preparation, inference-based sizing, and ExTuRBO for explainable, warm-started optimization, producing a final design report with Global and Elite sensitivity insights. Empirical results on PTM-based amplifiers show a 100% success rate and fast runtimes for the relaxed design space, while highlighting the importance of schematic information for tractable optimization in tighter design spaces. The approach offers a path toward explainable, efficient, and industry-ready automated analog sizing that couples numeric optimization with model-based reasoning and trust-enhancing analysis.

Abstract

Analog mixed-signal circuit sizing involves complex trade-offs within high-dimensional design spaces. Existing automatic analog circuit sizing approaches often underutilize circuit schematics and lack the explainability required for industry adoption. To tackle these challenges, we propose a Vision Language Model-optimized collaborative agent design workflow (VLM-CAD), which analyzes circuits, optimizes DC operating points, performs inference-based sizing and executes external sizing optimization. We integrate Image2Net to annotate circuit schematics and generate a structured JSON description for precise interpretation by Vision Language Models. Furthermore, we propose an Explainable Trust Region Bayesian Optimization method (ExTuRBO) that employs collaborative warm-starting from agent-generated seeds and offers dual-granularity sensitivity analysis for external sizing optimization, supporting a comprehensive final design report. Experiment results on amplifier sizing tasks using 180nm, 90nm, and 45nm Predictive Technology Models demonstrate that VLM-CAD effectively balances power and performance, achieving a 100% success rate in optimizing an amplifier with a complementary input and a class-AB output stage, while maintaining total runtime under 43 minutes across all experiments.

VLM-CAD: VLM-Optimized Collaborative Agent Design Workflow for Analog Circuit Sizing

TL;DR

VLM-CAD addresses the challenge of automatic analog circuit sizing by integrating a Vision-Language Model–driven collaborative agent workflow with schematic interpretation. It combines Image2Net for precise schematic annotation, a multi-agent circuit analysis pipeline, DC bias preparation, inference-based sizing, and ExTuRBO for explainable, warm-started optimization, producing a final design report with Global and Elite sensitivity insights. Empirical results on PTM-based amplifiers show a 100% success rate and fast runtimes for the relaxed design space, while highlighting the importance of schematic information for tractable optimization in tighter design spaces. The approach offers a path toward explainable, efficient, and industry-ready automated analog sizing that couples numeric optimization with model-based reasoning and trust-enhancing analysis.

Abstract

Analog mixed-signal circuit sizing involves complex trade-offs within high-dimensional design spaces. Existing automatic analog circuit sizing approaches often underutilize circuit schematics and lack the explainability required for industry adoption. To tackle these challenges, we propose a Vision Language Model-optimized collaborative agent design workflow (VLM-CAD), which analyzes circuits, optimizes DC operating points, performs inference-based sizing and executes external sizing optimization. We integrate Image2Net to annotate circuit schematics and generate a structured JSON description for precise interpretation by Vision Language Models. Furthermore, we propose an Explainable Trust Region Bayesian Optimization method (ExTuRBO) that employs collaborative warm-starting from agent-generated seeds and offers dual-granularity sensitivity analysis for external sizing optimization, supporting a comprehensive final design report. Experiment results on amplifier sizing tasks using 180nm, 90nm, and 45nm Predictive Technology Models demonstrate that VLM-CAD effectively balances power and performance, achieving a 100% success rate in optimizing an amplifier with a complementary input and a class-AB output stage, while maintaining total runtime under 43 minutes across all experiments.
Paper Structure (16 sections, 1 equation, 5 figures, 5 tables)

This paper contains 16 sections, 1 equation, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Machine Learning and Reinforcement Learning Methods
  4. LLM-based Methods and Agentic Workflows
  5. Circuit Schematics and VLM
  6. VLM-CAD Agentic Workflow
  7. Image2Net Interpretating Circuit Schematic
  8. Circuit Netlist and Specifications Analysis
  9. DC-OP Sizing
  10. Inference-Only Sizing
  11. ExTuRBO Sizing and Explaining
  12. Experiments
  13. Experimental Setup
  14. Results and Analysis
  15. Ablation Study
  16. ...and 1 more sections

Figures (5)

  • Figure 1: VLM-CAD overview. We include five stages for VLM-CAD: A. Image2Net Interpretating Circuit Schematic (see Sec. \ref{['text:Phase_A']}), B. Circuit Netlist and Specifications Analysis (see Sec. \ref{['text:Phase_B']}), C. DC-OP Sizing (see Sec. \ref{['text:Phase_C']}), D. Inference-Only Sizing (see Sec. \ref{['text:Phase_D']}) and E. Sizing and Explaining (see Sec. \ref{['text:Phase_E']}).
  • Figure 2: Image2Net for circuit schematic interpretation, including three stages: Component Detection, Corner Detection, and Graph Construction (see Sec. \ref{['text:Phase_A']}).
  • Figure 3: Schematics of the two amplifiers we use in our experiments: (a) an amplifier with complementary input and a class-AB output stage, (b) a two-stage Miller operational amplifier.
  • Figure 4: Final design report that includes Global Sensitivity, Elite Sensitivity, and explanations based on these quantified data.
  • Figure 5: Universal cost distribution of Phase 3 for the two-stage Miller operational amplifier, organized from highest to lowest costs.