Machine Learning in Routing
The idea
Everything a router does is steered by heuristics, and each one is a candidate for a learned replacement:
| lever | today, hand-tuned | learned proposal |
|---|---|---|
| congestion prediction | RUDY (Rectangular Uniform wire DensitY)-style analytic maps; trial global routing | CNN (Convolutional Neural Network) / GNN (Graph Neural Network) hotspot models from the placement (RouteNet lineage) |
| net ordering | criticality and fanout heuristics | learned ordering policies (RL (Reinforcement Learning)) |
| cost weights | congestion price schedules tuned per node | learned per-region weights |
| rip-up victims | overflow bookkeeping | learned victim scoring |
The first row is the most developed. A model maps placement features (cell density, pin density, net bounding boxes) to the congestion hotspot map, which is otherwise known only after routing. An accurate prediction lets the router price those cells from the first iteration rather than finding them through overflow, and lets placement improve routability before routing runs. The second use is where the approach is applied in practice.
The recipe (the prediction loop)
Step 1 (featurize). Rasterize the placement into feature channels (density, pins, net boxes) or build the netlist graph.
Step 2 (predict). The model outputs a hotspot map with no routing run at all.
Step 3 (route under the prediction). Candidates are priced with the predicted cells pre-charged, so routes steer around trouble before any overflow exists.
Step 4 (verify and retrain). Routing runs, and the actual congestion is compared with the prediction. Every completed run is labeled training data for the next model. The same low-cost verification that supplies these labels also sets the accuracy the model must exceed to be useful.
Why routing resists learning
Routing has a property placement lacks: its quality can be checked exactly and cheaply. Running the router and counting the resulting overflow and design-rule violations gives the true cost of a routing decision in minutes. The classical search is also near-optimal per net, so a learned component must outperform a baseline whose own result is quickly verifiable. Mispredictions carry a cost: an unpredicted hotspot leads routes into congestion, an over-predicted hotspot forces unnecessary detours, and the negotiation must run in both cases to correct the routing. Learned models therefore reduce the number of iterations the loop needs; none has replaced the loop.
Watch it work: predict, route, verify
One net on a 30 x 20 placement through the full loop: features, the predicted hotspot map, three candidates priced against it, the measured congestion after routing with correct predictions, missed hotspots, and false positives counted, and the cost of each prediction error. Every value is computed from the stated model.
Pros, cons, and the trade-off
Strengths
- Accurate prediction: placement-stage congestion models are accurate enough to guide routability-driven placement.
- Low labeling cost: every completed routing run yields labeled training data, so data collection is inexpensive.
- Room to improve: net ordering, cost weights, and rip-up selection remain hand-tuned, leaving scope for learned replacements.
Weaknesses
- Strong verifiable baseline: negotiated search is near-optimal per net and its cost is cheap to check, setting a high bar.
- Prediction error costs iterations: the negotiation must correct any routing the model steered incorrectly.
- Limited generalization: a model trained on one design family and process node loses accuracy on another, as the rule deck and pitches change.
The trade-off
Machine learning is applied to routing where the cost balance favors it: predicting expensive outcomes cheaply, before routing runs. Inside the router the classical search remains, because it can be verified more cheaply than a model can be trusted. The open direction is not replacing the search but determining, earlier and more accurately, what it is asked to route.
Is it actually used today?
The evidence for machine learning in routing splits by where the model sits.
- Before routing: documented in commercial tools. Production placers use machine-learning congestion and routability prediction, and the research line (RouteNet, the CircuitNet dataset) is active and reproducible.
- Inside the router: no production router documents a learned component in its search or negotiation; reported uses are limited to prediction and parameter tuning.
- Compared with placement: reinforcement-learning macro placement has been published as a production result (Nature, 2021); routing has no equivalent, for the verification reasons above.