DaVinci: Performance-Driven Analog Routing via Multimodality Guidance Prediction

Analog routing automation remains challenging because variability in parasitics and mismatch are highly design- and topology-dependent, limiting the transferability of heuristics across different circuits. Existing methods typically imitate manual routing heuristics or use only one type of routing feature, missing valuable information from both the layout and netlist. In this work, we introduce DaVinci, a unified framework that employs multi‑modal guidance prediction with neural networks to generate routing directives explicitly guided by performance objectives. By combining detailed layout cost maps with netlist connectivity using a heterogeneous message‑passing architecture and multi‑scale feature extraction, DaVinci incorporates performance gradients directly into the routing process. This integration enables the model to capture both local and global circuit interactions, aligning routing decisions with post-layout performance targets. Experimental results demonstrate that our method significantly outperforms conventional approaches, delivering enhanced routing quality and design efficiency.