SCALE targets local place-and-route design-rule violation repair at sub-2nm nodes. It serializes multi-layer BEOL geometry as structured text and self-supervises a language model by reconstructing randomly masked polygons. Natural-language constraints and high-temperature sampling then produce candidate layouts, which an industrial signoff DRC checker labels for training a domain-adapted DRC-VLM. According to the abstract, its geometric guidance improves existing repair agents’ solve rates by 12–25 percentage points, reaching up to 97% across 100 real cases involving enclosure, spacing, width, and color-spacing violations.
No heat snapshots are available in the last 24 hours.