One Flow-Matching Model, Two Routing Abstractions: Generative Routing for Verified Analog/Mixed-Signal Layout

Learning-based routers often predict a global routing topology or guidance map, leaving layer assignment, track legalization, and design-rule closure to a separate detailed routing engine. We investigate whether a single generative routing model can instead support both global routing and technology-aware detailed routing for analog/mixed-signal (AMS) layouts. AMSFlowRoute trains a conditional flow-matching model, a continuous-time generative approach that learns a routing velocity field, on a synthetic dataset and expert layouts. At inference, the same frozen field is combined with stage-specific guidance: Steiner and occupancy guidance for multi-net global routing, and process design kit (PDK) guidance for directional metal and via generation. A deterministic realization stage then snaps segments to legal tracks, inserts vias, resolves any remaining conflicts, and verifies the layout. On a common 600-net synthetic benchmark, AMSFlowRoute achieves 100\% connectivity with a 0.97$\times$ wirelength ratio, compared with 58.2\% connectivity and 2.59$\times$ wirelength for a diffusion baseline. On 11 example expert-generated AMS circuits comprising 159 signal nets, progressive routing reaches 100\% connectivity at 1.02$\times$ the rectilinear minimum spanning tree (MST) reference. The complete physical pipeline passes design-rule checking (DRC) and layout-versus-schematic (LVS) verification for all 11 circuits. These results show that the learned component can be shared across routing stages, while stage- and technology-specific requirements are handled by inference-time guidance and deterministic physical realization.