RouteDiff: Constraint-Aware Latent Diffusion Transformer for PCB Routing

Learning-based routing has been actively studied, but prior work has focused primarily on VLSI and integrated-circuit routing. PCB routing is similarly multimodal, admitting multiple valid solutions for a given placement and netlist, while additionally supporting a broad range of non-Manhattan and gridless geometries and imposing complex per-net design constraints such as track width and clearance. We therefore formulate PCB routing as a fixed-order, per-net conditional generation problem in which the routing geometry of a target net is generated from the current routing state and its design constraints. To this end, we propose RouteDiff a constraint-aware latent diffusion transformer. RouteDiff uses heterogeneous constraint conditioning to integrate spatial routing context derived from the current routing state and design constraints, net-level scalar design constraints, and target-pad geometry through distinct conditioning pathways to generate layer-aware routing geometry. We further introduce Geometry-Preserving Route Realization (GPRR), which converts generated routing images into physical trace geometry while preserving the predicted geometry, thereby reducing the extent to which downstream routing can obscure quality differences between generative models. On a public real-board PCB benchmark and industrial production-board data, RouteDiff consistently outperforms adapted generative and prediction-based baselines, achieving higher pre-GPRR F1 scores against the reference routing, higher post-GPRR routability, and fewer design-rule violations (DRVs). It also produces well-defined octilinear routing geometry, and its output responds consistently to changes in the track-width constraint. These results demonstrate the effectiveness of explicitly incorporating complex PCB design conditions into generative routing.