An Automated Debugging and Evolution Framework for Analog Code-to-Layout Generators

Analog code-to-layout generators make analog layout design more programmable, modular, and reusable, but debugging generated layouts remains a major bottleneck. Physical verification reports identify DRC/LVS failures and quality issues at the layout-artifact level, while effective repair requires edits to the generator source code, parameters, and hierarchy that produced the failing layout structures. This paper presents an automated debugging and evolution framework for analog code-to-layout generators. Instantiated on gLayout, this framework introduces a Source-Mapped Generator Runtime (SMGR) that links generated geometry back to source calls. This mapping enables a failure localizer to translate DRC/LVS failures and PEX-derived quality signals into structured bug contexts. The context is then passed to a repair agent trained from verifier-grounded fault-injection examples, which applies generator-code edits. A lightweight QoR surrogate further prioritizes repair candidates using QoR proxies, reducing unnecessary verification. Coupled with a layout coding agent, the framework forms an agentic workflow that iteratively generates, verifies, localizes, and repairs generator code to improve layout correctness and quality. In end-to-end benchmarks, full local loop reaches 99.16\%/98.21\% Clean@1h on two prompt sets. Relative to the strongest proprietary coding-agent raw-log baseline, it reduces time-to-clean by 3.0$\times$/4.4$\times$ and normalized logged text-token volume by 18.4$\times$/18.3$\times$; used as source-localized feedback, it improves paired proprietary-agent baselines by 11.0/12.8 Clean@1h percentage points on average.