acceptodds
Under review as a conference paper at ICLR 2027

RouteDiff: A Constraint-Aware Diffusion for PCB Routing with Geometry-Preserving Route Realization

Abstract

Recent learning-based routing methods have focused primarily on VLSI and IC routing, while learned PCB routing remains comparatively underexplored. Printed circuit board (PCB) routing similarly admits multiple valid solutions for a given placement and netlist, with routing geometries that include non-orthogonal segments and are subject to per-net rules such as track width and clearance. We consider PCB routing as sequential conditional generation, where each target-net route is conditioned on the current routing state, target-pad geometry, explicit per-net design rules, and information from the remaining netlist. To model these conditions, we propose RouteDiff, a constraint-aware latent diffusion transformer that integrates spatial routing conditions, scalar design rules, and target-pad geometry through heterogeneous conditioning pathways. We further introduce Geometry-Preserving Route Realization (GPRR), a realization protocol that converts generated routing images into physical PCB traces while limiting modification of the generated geometry. On a public PCB benchmark and an industrial PCB dataset, RouteDiff achieves the highest F1 to the reference routes and, after GPRR realization, the highest routability and the fewest design-rule violations among the learned baselines, while responding distinctly to width and clearance interventions. Our evaluations support constraint-aware generation and geometry-preserving realization as complementary components for learned PCB routing.

Primary area:
generative models
Keywords:
pcb routing, chip design, diffusion, dit, generative model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KBMLOpenReview ↗
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
AuthorsOpenReview ↗

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.

Training fractionBoardsNetsF1Routability (%)
10%491,7050.488651.6 ± 0.9
25%1223,5380.479159.1 ± 0.4
50%2446,8770.539160.4 ± 0.3
100%48714,4710.572067.1 ± 0.8

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.

ModelZero-shot routability (%)Fine-tuned routability (%)Zero-shot / fine-tuned DRVs
RouteDiff48.4 ± 0.353.0 ± 0.51,299 / 744
UNet48.850.81,298 / 1,203
cGAN-P38.037.01,823 / 1,565
cGAN-H36.738.02,016 / 1,566

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.

Reviewer_KBMLOpenReview ↗

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.
AuthorsOpenReview ↗

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 / netsRef., full GPRRRef., vect. + PNSRouteDiff, full GPRRPNS: unguided → RouteDiff
<13 / 31592.598.335.167.7 → 82.1
1 to <219 / 73099.399.472.385.9 → 96.4
2 to <428 / 70999.799.267.979.7 → 93.6
≥417 / 21499.799.973.987.2 → 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.

Reviewer_zwtGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work proposes a two-stage method to implement PCB routing. The first stage is the RouteDiff, a constraint-aware latent diffusion transformer that integrates spatial routing conditions, scalar design rules, and target pad geometry through heterogeneous conditioning pathways. Then, a Geometry-Preserving Route Realization protocol is introduced to convert routing images into physical PCB traces. The results show that the proposed method can achieve the best performance for public PCB benchmarks and industrial PCBs.

Critical strengths
  1. The problem is important, and the method is well motivated. PCB routing is difficult because routes must connect the correct pads while following width and clearance rules. The paper considers these constraints directly during generation. Using separate inputs for spatial information, design rules, and pad geometry is a reasonable approach.

  2. The author compared the proposed method with a bunch of other methods, and its performance is stronger than that of any other method.

Critical weaknesses
  1. The proposed framework mainly combines existing techniques, including latent diffusion, DiT, and conditional embeddings. What is the main novelty from a machine learning perspective? I would like the authors to explain what new learning insights this work provides beyond applying existing models to PCB routing.

  2. Although RouteDiff improves over the evaluated learning-based baselines, its routability is still only 69% on the public dataset and 53.0% on the industrial dataset. These results seem far from what is needed for practical PCB routing. How would this method benefit real PCB design workflows? A comparison with a strong traditional PCB router would help clarify its practical value.

  3. RouteDiff generates routes one net at a time and fixes the previous routes. However, an early route may block the paths needed by later nets. How sensitive is RouteDiff to the routing order? Can it handle cases where earlier routes need to be changed to complete the whole PCB? Is it possible to trace back to revise, or not?

Questions for authors

For the sequential routing method, I was wondering how to handle and evaluate the violations.

2 comments
AuthorsOpenReview ↗

Response to Reviewer zwtG

We thank the reviewer for recognizing the importance of the task and the motivation for separating spatial conditions, design rules, and pad geometry.

W1. Novelty from a machine learning perspective.

The work provides evidence about learning and evaluating physical routing geometry (G4). The interventions show distinct effects of scalar and spatial conditioning on width and clearance. With conditioning fixed, changing v-prediction to ε-prediction reduces routability from 69.2% to 63.8%. The additional controls show that a small RouteDiff retains an advantage over cGAN-P even after the latter receives the additional conditions.

The evaluation provides a further insight: downstream search can improve weak predictions while obscuring differences between generators. GPRR limits this replacement, while the new common-engine CAD evaluation separately establishes practical utility. Our contribution lies in these supported design and evaluation choices rather than a new attention or diffusion block.

W2. Practical value and comparison with traditional routers.

The reported 69.2% and 53.0% are routability after GPRR and conflict-based evaluation, without board-level rip-up and reroute. They measure generated routes under a shared realization protocol, rather than completion by a full routing system.

We therefore added a separate CAD evaluation using generated waypoints as guidance for KiCad’s PNS engine (G3; §5.5). Under the same engine, RouteDiff raises the board-mean connection rate from 82.83% without guidance to 94.31% with one sample; other learned guides reach 84.09–85.52%. At K=8, RouteDiff reaches 96.57% connection rate and 62.2% complete boards, compared with 94.26% and 62.2% for Freerouting p=10. At K=32, the rates rise to 97.37% and 66.7%.

This establishes a concrete workflow benefit: learned geometry improves the engine’s ability to connect nets under the evaluated rules. It comes at higher runtime and with longer routes and more vias than Freerouting, which we report explicitly.

W3. Routing order and revising earlier routes.

We evaluated the same model under permuted orders on the D3 test set, using three sampling seeds. AR denotes autoregressive generation conditioned on preceding generated routes; RP denotes reference-prefix conditioning. Intervals are board-level bootstrap 95% confidence intervals for paired changes.

OrderAR routability (%)ΔARRP routability (%)ΔRP
Paper: small pad bbox first75.5—69.3—
Random 174.7−0.8 [−2.3, +0.6]67.7−1.6 [−3.0, −0.1]
Random 275.5−0.1 [−1.7, +1.6]67.9−1.3 [−2.8, +0.2]
Random 374.5−1.1 [−2.4, +0.2]67.6−1.7 [−3.1, −0.3]
Reverse: large first72.4−3.1 [−4.9, −1.3]68.1−1.1 [−3.1, +0.9]

The paper-order rows were rerun on a different GPU type, giving 75.5/69.3 versus the manuscript’s 75.2/69.2; all changes use the rerun as their paired reference. Unrounded values were used to calculate changes.

Random orders have modest effects in AR, with all three intervals including zero. Reversing the order causes a larger, statistically resolved decrease of 3.1 points. RP also shows some sensitivity, including decreases whose intervals exclude zero. The results therefore support modest sensitivity to the tested random orders, not order invariance.

RouteDiff itself fixes earlier routes and does not backtrack. We now state this explicitly in Appendix A.1. Its outputs can nevertheless guide an engine that revises routes. With PathFinder’s negotiated-congestion rerouting, RouteDiff guidance increases routability from 74.1% to 85.5% and reduces DRVs from 17.9 to 14.3 per board (App. C, Table 17). In the PNS evaluation, unconnected nets are regenerated using the copper actually placed; this is distinct from learned backtracking.

Q. Handling and evaluating violations.

During generation, preceding routes enter the routing context, while the clearance-aware field supplies spatial information about obstacles and rules. This conditions the model to avoid conflicts but does not guarantee compliance.

Under GPRR evaluation, all routes are realized before board-level DRC. DRVs include violations before any removal. Routability is then computed by building a net-conflict graph and greedily removing conflicting nets until no violations remain; disconnected nets and nets violating fixed pads, existing copper, or the board boundary are excluded (App. A.2). Earlier decisions that obstruct later routes therefore affect the final score.

The CAD evaluation uses a separate operational criterion: counted nets must be connected and introduce no new short or clearance violation under the evaluated KiCad checks. This distinguishes diagnostic evaluation of generated geometry from routing with physical conflict handling.

AuthorsOpenReview ↗

Response to Reviewer zwtG

Dear Reviewer zwtG,

Thank you again for your review. As the discussion period ends on November 18 (AoE), we would appreciate your feedback on whether our response addresses your concerns. In particular:

  • ML contribution (W1): G4 states three findings supported by the experiments: the conditioning pathways have distinct geometric effects, the training formulation matters with conditioning held fixed, and downstream search can obscure differences between generators.
  • Practical value (W2): Under the same KiCad PNS engine, RouteDiff guidance raises the connection rate from 82.83% to 94.31% with one sample. With up to eight samples per net (K=8), it reaches 96.57%, compared with 94.26% for Freerouting p=10, with the same complete-board rate of 62.2% (§5.5).
  • Sequential routing and violations (W3/Q): The tested random orders change AR routability by at most 1.1 points, while reversing the order reduces it by 3.1 points. RouteDiff itself does not backtrack, and conflicts caused by earlier routes are counted in the final score. RouteDiff guidance raises PathFinder's routability from 74.1% without learned guidance to 85.5% (App. C).

Please let us know if any concerns remain so that we can address them before the discussion closes. If these results resolve your concerns, we would appreciate your reconsideration of the assessment.

Best regards, The Authors

Reviewer_j6ngOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a constraint-aware latent diffusion model for PCB routing. The method integrates spatial routing informationand target-pad geometry through heterogeneous conditioning. The authors also introduce Geometry-Preserving Route Realization to convert generated routing images into physical PCB traces while preserving the generated geometry. Experiments on PCB datasets demonstrate improvements over several baselines.

Critical strengths

The proposed heterogeneous conditioning mechanism integrates spatial routing information, scalar design rules, and target-pad geometry in a reasonable and effective way.

The paper addresses PCB routing with explicit geometric and design-rule constraints, the combination of heterogeneous conditioning and geometry-preserving route realization is reasonable and practically useful.

RouteDiff achieves better routing similarity, routability, and fewer design-rule violations than the evaluated learning-based baselines. The experiments also include industrial PCB data and autoregressive routing evaluations.

Critical weaknesses

RouteDiff has approximately 613M parameters, substantially more than most baselines. Moreover, the baselines do not receive the same design-rule and pad-geometry information. It is therefore difficult to determine how much of the improvement comes from the proposed method rather than increased model capacity and richer conditioning.

Although the method is described as constraint-aware, the generated routes still contain many design-rule violations. Interestingly, removing the scalar design-rule conditioning does not worsen routability or DRVs. This raises questions about how effectively the proposed conditioning improves physical constraint satisfaction.

The proposed method mainly combines existing latent diffusion architectures with standard conditioning mechanisms, including spatial feature concatenation and cross-attention. GPRR also relies largely on conventional image processing.

Questions for authors

Can the authors provide comparisons with capacity-matched diffusion baselines using the same conditioning information?

How does RouteDiff compare with conventional PCB routing algorithms in terms of complete routing success, wirelength, and design-rule compliance?

How well does the method scale to larger multilayer PCBs with more complex constraints?

2 comments
AuthorsOpenReview ↗

Response to Reviewer j6ng

We thank the reviewer for recognizing the value of heterogeneous conditioning and geometry-preserving realization. Your questions motivated additional controls and a direct evaluation with conventional PCB routers.

W1 / Q1. Capacity-matched diffusion baselines with the same conditioning.

We added two complementary controls (§5.3, Table 4; G1). RouteDiff-tiny retains RouteDiff’s conditioning at 11.2M parameters and achieves 64.6% routability, compared with 55.3% for the 11.4M cGAN-P. Conversely, augmenting cGAN-P with RouteDiff’s spatial conditions raises its F1 from 0.496 to 0.566 but its routability only to 58.0%. Adding scalar rules and pad tokens gives 0.561 F1 and 55.7% routability at 12.5M parameters.

These results test whether the gains require a large RouteDiff model and whether providing the additional inputs closes the gap for the strongest original baseline. Within RouteDiff, the fixed-conditioning comparison of v- and ε-prediction additionally isolates the training formulation: routability decreases from 69.2% to 63.8% with ε-prediction (Table 1). We distinguish these controls from a complete isolation of architecture, optimization, and training compute.

W2. Scalar-rule conditioning and remaining DRVs.

The scalar ablation retains the spatial field, which also encodes width and clearance. Its result therefore does not imply that removing all rule information leaves performance unchanged.

The direct interventions show different roles for the two pathways. A 3× width instruction changes generated width by 2.13× through scalar conditioning and 1.16× through the spatial field. GPRR redraws centerlines at the specified width, so this generated-width response need not yield a corresponding improvement in realized routability. For clearance, the spatial field restores compliance for 46.7% of the 75 selected nets with binding and attainable doubled-clearance targets, versus 10.7% for the scalar pathway alone (App. G).

We also investigated inter-net conflicts. Holding target predictions fixed, replacing preceding generated copper with the reference copper used for conditioning reduces conflicts from 47.2 to 18.1 per board. Thus, assembling independently generated routes under reference prefixes contributes substantially to the remaining conflicts. All such violations remain counted; G3 evaluates the complementary use of a routing engine to handle them.

W3. Novelty.

G4 states the learning contribution and supporting evidence. The contribution is the conditioning and training design for physical routing geometry, together with an evaluation that limits downstream replacement of model predictions. The pathway interventions, training-formulation ablation, and common-engine routing results provide evidence for these choices.

Q2. Conventional routers: completion, wirelength, and compliance.

The new CAD evaluation uses 67 public D3 test boards (§5.5; App. D).

RouteDiff + PNS achieves a 96.57% board-mean connection rate at K=8, compared with 94.26% for Freerouting p=10 and 90.69% for KRT. Its complete-board rate is 62.2%, matching Freerouting p=10. At K=32, these rates increase to 97.37% and 66.7%. Every counted connection introduces no new short or clearance violation under the evaluated KiCad checks.

On the 1,098 nets completed by every method and run included in Table 19, RouteDiff’s wirelength is close to the reference (ratios 0.992–0.995). Freerouting and KRT produce shorter routes, approximately 0.94 and 0.92 of reference length. RouteDiff also uses more vias and takes longer: at K=8, 47.9 versus 6.8 seconds per board for Freerouting p=10. Thus, the new results establish a completion benefit from learned guidance while identifying its geometric and computational costs.

Q3. Larger multilayer boards and complex constraints.

The industrial test boards contain an average of 155.8 target nets and 877.7 pads, versus 29.4 and 126.5 on the public benchmark (Tables 8–9). RouteDiff achieves the highest routability among the evaluated learned methods on these boards (53.0% versus 50.8%) and fewer DRVs (744 versus 1,203 per board; Table 6).

This demonstrates adaptation to larger two-layer boards, not multilayer scalability. The fixed raster representation also incurs losses: reference routes achieve 84.1% routability after the same rasterize–realize round trip (§5.6). Additional layers require changes to the output representation and realization. G5 and Appendix A.1 distinguish conditions compatible with the existing inputs from capabilities that require further development.

AuthorsOpenReview ↗

Response to Reviewer j6ng

Dear Reviewer j6ng,

Thank you again for your review. As the discussion period ends on November 18 (AoE), we would appreciate your feedback on whether our response addresses your concerns. The main additions are:

  • Capacity and conditioning (W1/Q1): RouteDiff-tiny achieves 64.6% routability at 11.2M parameters. At comparable sizes, cGAN-P reaches 55.3% originally, 58.0% with RouteDiff's spatial conditions, and 55.7% with additional rule scalars and pad tokens (§5.3, Table 4).
  • Rule conditioning (W2): The scalar ablation retains rule information in the spatial field. The interventions show distinct effects of the two pathways on generated width and clearance compliance, and the conflict analysis shows that assembling routes generated under reference prefixes contributes substantially to the remaining conflicts (G2, App. G).
  • Conventional routers (Q2): With RouteDiff guidance and up to eight generated samples per net (K=8), KiCad PNS achieves a 96.57% connection rate versus 94.26% for Freerouting p=10, with the same complete-board rate of 62.2%. Runtime, wirelength and via costs are reported in §5.5 and App. D.

Please let us know if any concerns remain so that we can address them before the discussion closes. If these results resolve your concerns, we would appreciate your reconsideration of the assessment.

Best regards, The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response (1/2)

We thank the reviewers for their constructive feedback. Reviewers found the separate conditioning pathways reasonable (j6ng, zwtG) and appropriate (KBML), and noted gains over learned baselines (j6ng, zwtG) and the intervention evidence (KBML). All changes in the revised PDF are marked in blue. We have added capacity and conditioning controls and comparisons with conventional routers to the paper, and report data-scaling, industrial-transfer, and routing-order analyses in this response. We address shared concerns below and provide reviewer-specific responses separately.

G1. Model capacity and conditioning (j6ng, KBML).

The additional controls address model size and conditioning access separately. RouteDiff-tiny retains the full conditioning design while reducing the model to 11.2M parameters. We also augment cGAN-P with RouteDiff’s spatial conditions, then additionally supply the rule scalars through FiLM and pad tokens through cross-attention (§5.3, Table 4).

ModelParamsF1Routability (%)
cGAN-P11.4M0.49655.3
+ spatial conditions11.4M0.56658.0
+ scalar rules and pad tokens12.5M0.56155.7
RouteDiff-tiny11.2M0.57164.6
RouteDiff613.1M0.59769.2

The cGAN-P additions are cumulative. Its spatial inputs contain the routing context, clearance field, and remaining-net demand used by RouteDiff.

At comparable parameter counts, RouteDiff-tiny exceeds the original cGAN-P by 9.3 percentage points in routability. Providing spatial conditions substantially improves cGAN-P’s F1, bringing it close to RouteDiff-tiny (0.566 versus 0.571), but leaves a 6.6-point routability gap. Adding scalar rules and pad tokens does not close this gap under the tested training recipes. These controls show that the observed advantage does not require the 613M model and is not reproduced simply by supplying the additional inputs to cGAN-P.

G2. Rule conditioning and remaining violations (j6ng, zwtG).

Removing the scalar pathway does not remove all rule information: the clearance-aware spatial field also encodes width and clearance. The ablation therefore measures the additional contribution of scalar inputs. Removing them lowers F1 from 0.597 to 0.580, but does not worsen aggregate routability or DRVs (Table 24).

The intervention experiments clarify this result. A 3× width instruction produces a 2.13× change in generated width through the scalar pathway, versus 1.16× through the spatial field (App. G.2). GPRR subsequently redraws extracted centerlines at the specified track width, limiting how directly improved generated width translates into realized routability or DRVs. For clearance, among 75 nets with binding and attainable doubled-clearance targets, the spatial field restores compliance for 46.7%, versus 10.7% for scalar conditioning alone (App. G.3).

We also examined the remaining conflicts. In the reference-prefix setting, each net is conditioned on preceding reference routes, while evaluation assembles the generated routes. Holding the generated target routes fixed, conflicts with preceding routes decrease from 47.2 per board against generated copper to 18.1 against the corresponding reference copper. This supports a substantial contribution from the mismatch between conditioning context and assembled routes; these conflicts remain included in the reported DRVs.

G3. Practical routing utility and conventional routers (j6ng, zwtG).

Under the same KiCad push-and-shove routing engine (PNS), RouteDiff guidance raises the board-mean connection rate from 82.83% without guidance to 94.31% with one sample. Other learned guides reach 84.09–85.52%. This evaluation uses extracted waypoints without GPRR’s local bridging or full A* rerouting (§5.5; App. D).

MethodConnection rate (%)Complete boards (%)Time (s/board)
Freerouting, p=593.0656.25.8
Freerouting, p=1094.2662.26.8
Freerouting, p=3093.8059.210.5
KRT90.6958.212.3
PNS, unguided82.8325.43.1
PNS + RouteDiff, K=194.3150.729.3
PNS + RouteDiff, K=896.5762.247.9
PNS + RouteDiff, K=3297.3766.784.9

Here, p is the Freerouting stop pass and K is the maximum number of generated samples per net. Complete boards have all evaluated target nets connected. Counted connections introduce no new short or clearance violation under the evaluated KiCad checks.

At K=8, RouteDiff exceeds Freerouting p=10 in connection rate and matches its complete-board rate; K=32 improves both. RouteDiff with one sample also exceeds the strongest alternative learned guide with 32 samples (94.31% versus 93.18%; Table 18). These gains incur higher runtime, and Freerouting produces shorter routes with fewer vias. The practical contribution is effective learned guidance for a CAD router, with the trade-offs reported explicitly.

AuthorsOpenReview ↗

General Response (2/2)

G4. Contribution from a machine learning perspective (j6ng, zwtG).

Our contribution is an empirically supported design and evaluation framework for generating physical routing geometry under heterogeneous conditions. The experiments provide three findings.

First, the conditioning pathways have distinct effects. Scalar inputs strongly control generated track width, whereas spatial clearance fields more effectively restore clearance compliance in the intervention study (G2). Removing individual pathways reduces reference F1, and retaining only the routing context gives 0.536 F1 and 62.8% routability (Table 24). These results support the combined design within RouteDiff and identify the different roles of its conditioning pathways.

Second, the training formulation matters even with the conditioning held fixed. Replacing v-prediction with ε-prediction reduces F1 from 0.597 to 0.528 and routability from 69.2% to 63.8% (Table 1). Together with the small-model and enriched-baseline controls in G1, this provides evidence beyond applying a larger off-the-shelf generator with additional inputs.

Third, downstream realization affects what can be concluded about a generator. Full A* realization can replace substantial portions of a prediction, improving weak outputs and compressing differences between models. GPRR retains more of the generated geometry, allowing reference agreement and physical routing quality to be assessed together (§5.4, Fig. 3). Its contribution is this evaluation protocol, rather than novelty in thinning or morphology. The common-engine PNS comparison additionally tests whether the learned geometry remains useful when a CAD router handles physical routing.

The individual architectural components are established. The new evidence concerns how they should be combined and evaluated for this task: different condition forms influence different geometric properties, reference F1 alone does not capture routability, and the choice of realization can obscure differences between generators. We have clarified this focus in the Introduction.

G5. Evaluated scope and extensibility (j6ng, KBML).

We evaluate two-layer trace routing with explicit width and clearance rules. Our question is whether a generator conditioned on physical PCB inputs can learn routing geometry that is useful downstream, and this question can be studied without solving full production routing at once. Prior PCB work similarly studies parts of the design problem in isolation. U-Net-A* defines its learned routing problem on two-layer PCB layouts (Yin et al., 2023). FanoutNet automates PCB fanout, a stage performed before routing (Li et al., AAAI 2023). Fang et al. (DAC 2024) study length matching on existing traces, including differential pairs, illustrating how additional routing requirements can be addressed as focused optimization problems. Appendix A.1 now makes the boundary between our task and full production routing explicit.

The public targets include 18,230 of 18,864 multi-pad nets (96.6%) across 618 boards. Power and differential nets are included as individual nets; this does not evaluate power-distribution requirements or coupled differential-pair constraints. Ground nets and nets without reference copper are excluded. The industrial targets are digital signal nets.

Keep-out regions and layer preferences could be supplied through the existing spatial and per-net inputs, although compliance would require appropriate training and evaluation. More copper layers, coupled differential pairs, length matching, impedance, and via restrictions require changes to the output, realization, or objectives.

Within this scope, RouteDiff’s geometry improves routing under both a shared realization protocol and a CAD engine that enforces the evaluated physical rules.

References for G5.

  • Yin et al. Unet-Astar: A Deep Learning-Based Fast Routing Algorithm for Unified PCB Routing. IEEE Access, 2023.
  • Li et al. FanoutNet: A Neuralized PCB Fanout Automation Method Using Deep Reinforcement Learning. AAAI, 2023.
  • Fang et al. Obstacle-Aware Length-Matching Routing for Any-Direction Traces in Printed Circuit Board. DAC, 2024 (arXiv:2407.19195).

Dear Reviewers,

The authors have posted their rebuttals. As the author-reviewer discussion period is coming to a close soon, please read their responses and engage in the discussion. Thank you for your time and effort.

Best,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.