- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes RouteDiff, a constraint-aware latent diffusion transformer for sequential, two-layer PCB routing. The model incorporates spatial routing context, per-net width and clearance rules, and variable-length pad geometry through heterogeneous conditioning pathways. The paper also introduces Geometry-Preserving Route Realization (GPRR), which converts generated images into physical traces while attempting to preserve the predicted geometry, and studies a learned waypoint decoder, WAND. Experiments on public and industrial PCB datasets demonstrate improvements over the evaluated learned baselines.
Critical strengths
Learned routing is a meaningful application with challenges that differ from conventional global routing. Explicitly representing per-net rules and layer-aware pad geometry is well motivated. The separation of spatial, scalar, and token-based conditions is appropriate for their different roles. The width and clearance intervention experiments provide useful evidence that the model responds to these inputs rather than merely reproducing reference patterns.
Critical weaknesses
Real-world PCB routing involves multilayer boards, differential pairs, length matching, impedance requirements, power/ground routing, via constraints, and other manufacturing and electrical considerations. The current evaluation is restricted to two-layer routing with explicit width and clearance constraints, with ground nets excluded and industrial targets limited to digital signal nets. The method is trained on 487 public boards containing 14471 target nets, followed by industrial fine-tuning on 227 boards containing 47761 target nets. However, the paper provides no data-scaling or low-data analysis. GPRR preserves generated geometry more effectively than full A* realization, but it is not search-free. For RouteDiff, 18.5% of nets require local bridging and 7.5% require full A* rerouting. Moreover, connectivity recovery is permissive and does not treat existing copper as a hard obstacle. RouteDiff differs from the baselines in model capacity, output resolution, training budget, and available conditioning. In particular, the full model has approximately 613 million trainable parameters, while several route-generating baselines have approximately 11–43 million parameters. The baselines also lack the explicit rule scalars and variable-length pad tokens supplied to RouteDiff.
Questions for authors
How does performance change when training on smaller fractions of the public training boards? What is the zero-shot performance on industrial boards before fine-tuning? Which production-level routing constraints can the current architecture accommodate without redesigning its output representation? Can a baseline supplied with the same rule and pad conditioning narrow the observed performance gap?
6 comments
Benchmark diversity and method limitation
Dear authors, Thank you for your reply. However, I have several critical concerns about the methods and experimental results.
- The proposed methods only handle two-layer boards. The two-layer board design is easy to handle with EDA tools or by humans. Thus, is it necessary to need this method?
- According to the method, it is difficult to extend the method to more than two-layer board designs. Because the method relies on pad location. In the inner layer, pad locations are missing, and the search space is enlarged. It is difficult to handle these cases.
- The performance relies on the generated routing image resolution. If there is a huge mismatch between routing image resolution and trace width, the generated routing image guide will generate disturbance for GPRR.
- In addition, I am wondering how much diversity there is between the training and testing datasets. According to my understanding, the learning method heavily relies on similarity between the training and testing datasets.
- Based on the last concern, if one testing layout is different from the training layout but the routed-90 testing layout is similar to the training layout, can you handle this case using the proposed method? I am wondering whether the proposed method can handle rotational similarity and symmetric similarity, since the performance relies on the similarity between the training and testing datasets.
Response to follow-up concerns
Thank you for the follow-up. We respectfully disagree that restricting the evaluation to two-layer boards makes the learning problem unnecessary. We separate this scope question from the technical concerns about resolution and generalization examined below.
Our previous response provided capacity and conditioning controls, the requested data-scaling and zero-shot analyses, and a CAD-engine evaluation without GPRR recovery. We would appreciate clarification on which original concerns remain after these results.
1. Two-layer boards. Our question is whether a generator conditioned on physical PCB inputs learns geometry that improves routing. Two-layer boards are not uniformly easy. On the 67 test boards, Freerouting completes all target nets on at most 62.2% of boards and KRT on 58.2% (Table 5). In KiCad PNS, RouteDiff guidance raises the connection rate from 82.83% to 94.31%, and to 96.57% at K=8 (Freerouting: at most 94.26%).
2. Inner layers. Pad conditioning does not inherently restrict the formulation to two layers. Layers are represented by separate routing-image channels and layer embeddings in pad tokens. This provides a basis for extension through additional layer encodings, with adapted realization and multilayer training data. Missing surface pads on inner layers do not remove terminal information. Inner-layer routes reach surface or through-hole pads through vias, and pad tokens identify terminals rather than requiring a pad on every layer. We have not demonstrated this extension, but the two-layer setting suffices for the question studied here (point 1).
3. Resolution. We agree that a fixed 512² raster limits sub-pixel geometry, as the paper reports for the industrial boards (84.1% reference routability). To separate representation from model, we realized reference images, grouped by median rule width, along two paths: full GPRR (with local bridging and A* recovery where needed, routability), and GPRR’s vectorization followed by KiCad PNS without GPRR’s local bridging or full A* rerouting (connection rate, Table 5). Values are board means (%), with reference-prefix generation for GPRR and K=1 for PNS.
| Width (px) | Boards / nets | Ref., full GPRR | Ref., vect. + PNS | RouteDiff, full GPRR | PNS: unguided → RouteDiff |
|---|---|---|---|---|---|
| <1 | 3 / 315 | 92.5 | 98.3 | 35.1 | 67.7 → 82.1 |
| 1 to <2 | 19 / 730 | 99.3 | 99.4 | 72.3 | 85.9 → 96.4 |
| 2 to <4 | 28 / 709 | 99.7 | 99.2 | 67.9 | 79.7 → 93.6 |
| ≥4 | 17 / 214 | 99.7 | 99.9 | 73.9 | 87.2 → 95.3 |
Over all test boards, reference images reach 99.26% routability with full GPRR and 99.42% connection rate with vectorization followed by PNS. For the 64 boards with median rule width of at least one pixel, group rates range from 99.2% to 99.9%. The raster thus retains sufficient information for high routing success. On the three sub-pixel boards, also the largest (width and board size are confounded), reference routability with GPRR falls to 92.5%, confirming a representation-and-realization limitation. RouteDiff reaches 35.1% with GPRR, yet its guidance raises PNS connection from 67.7% to 82.1%. Learned guidance improves PNS in every width group, and the remaining gap motivates better generation and representation.
4. Train-test similarity. The split assigns whole boards and removes near-duplicates (12 boards at fingerprint cosine ≥ 0.95). We compared each test board with every training board under all eight rotations and reflections, using 32×32 per-layer copper and pad-placement fingerprints. Test and validation boards are similarly close to the training set, with medians 0.319 vs. 0.326 (copper) and 0.280 vs. 0.269 (placement). The closest test board (0.628) is a variant of a training design with different placement (0.344). The 24 test boards from repositories shared with training boards reach 67.7% versus 70.0% for the rest (difference 95% CI [−11.4, 7.0]). After controlling for net count, area and pad density, we find no evidence of a positive association between routability and copper similarity (partial Spearman −0.02, 95% CI [−0.27, 0.19]).
5. Rotation and reflection. Training applies the same D₄ transform (four rotations × reflection) to the target, spatial conditions and pad tokens (App. A.2), and the audit above covers these transformations. We also applied each of the seven non-identity transforms to the inputs of every test net (same checkpoint and seed, outputs mapped back). Routability ranges from 67.6% to 70.4%, against 68.6% to 69.6% over three seeds without transformation. One reflection is 2.0 points lower (paired 95% CI [−3.9, −0.2]), and F1 is about 0.01 lower. Performance is broadly stable across orientations, with measurable degradation under some transforms.
We would be grateful to know which central claim you consider still insufficiently supported, so that we can continue this discussion constructively. Thank you again for your time.
Response to Reviewer KBML
We thank the reviewer for the detailed assessment and for recognizing the value of explicit rule and pad representations and the intervention experiments.
Data scaling (Q1).
We trained the 48.7M RouteDiff on nested board-level subsets, keeping the number of optimizer steps and learning-rate schedule fixed. Each checkpoint was evaluated with three sampling seeds; the 100% row is the 48.7M model in Table 10.
Routability increases across these subsets. With 122 boards, it reaches 59.1%, exceeding the original cGAN-P baseline trained on all boards (55.3%). This is a reference comparison, not a capacity-matched data-efficiency claim. F1 is not monotonic between 10% and 25%, and each fraction uses one training run; the reported variation covers sampling seeds rather than training variability.
Zero-shot industrial transfer (Q2).
We evaluated public-data checkpoints on the 15 industrial test boards before fine-tuning, using the same evaluation protocol.
Without fine-tuning, RouteDiff performs similarly to UNet and above the two cGANs. Fine-tuning improves RouteDiff by 4.6 percentage points and reduces its DRVs by approximately 43%, yielding the strongest fine-tuned results among these methods. These results demonstrate useful transfer while showing that industrial adaptation remains beneficial.
GPRR recovery and permissiveness (W).
For RouteDiff, 74.0% of nets are connected without search, 18.5% require local bridging, and 7.5% require full A* rerouting (Table 20). Recovery can introduce conflicts because existing copper is not a hard obstacle at this stage. Those violations are retained in DRVs and affect routability under the subsequent conflict-removal protocol. Thus, recovery establishes connectivity without automatically establishing valid routing.
More directly, the new PNS experiment uses generated waypoints without GPRR’s local bridging or full A* rerouting. Under the same engine, RouteDiff raises connection rate from 82.83% without guidance to 94.31%, whereas the other learned guides reach 84.09–85.52% with one sample (App. D, Table 18). This isolates the value of learned guidance within that engine.
GPRR serves a complementary measurement purpose. Compared with full A* realization, it retains more of the prediction and reduces the extent to which search replaces model-specific geometry (§5.4, Fig. 3). We distinguish this diagnostic protocol from the CAD evaluation of practical routing utility.
Capacity, conditioning, resolution, and training budget (W, Q4).
The new controls are in §5.3, Table 4 (G1). RouteDiff-tiny achieves 64.6% routability at 11.2M parameters. At comparable sizes, cGAN-P achieves 55.3% originally, 58.0% with RouteDiff’s spatial conditions, and 55.7% with additional scalar-rule and pad-token pathways. The spatial inputs substantially improve reference F1 and narrow the routability gap to RouteDiff-tiny from 9.3 to 6.6 points; the rule and pad pathways do not narrow it further (8.9 points).
Rescoring all predictions on a common 64×64 grid also retains RouteDiff’s F1 advantage (0.581 versus 0.494 for cGAN-P; App. B, Table 13). This controls scoring resolution, not the models’ native output resolution. Training budgets remain method-specific and are listed in Table 12. These analyses separately examine model size, access to conditioning, and scoring resolution.
Production constraints and representational scope (W, Q3).
G5 explains the research scope and its precedents in PCB routing. We evaluate two-layer routing with width and clearance rules, covering 96.6% of public multi-pad nets. Power and differential nets are routed individually; ground nets are excluded, and industrial targets are digital signal nets.
Without redesigning the two-layer output, spatial keep-outs and per-net layer preferences could be supplied through the existing conditioning pathways, with appropriate training and validation. Additional copper layers require output and realization changes. Differential-pair coupling, length matching, impedance, and specialized via constraints require further mechanisms or objectives and are not demonstrated here. Appendix A.1 clarifies this scope. Within it, the new CAD results establish that RouteDiff’s geometry improves an existing routing engine under the evaluated physical checks.