MAGNet: A Multi-Scale Attention-Guided Graph Fusion Network for DRC Violation Detection

Hong Cai Chen; Weihan Lu

REVIEW 5 major objections 7 minor 32 references

MAGNet: A Multi-Scale Attention-Guided Graph Fusion Network for DRC Violation Detection

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MAGNet claims that fusing a multi-scale, attention-guided U-Net with a tile-based graph neural network predicts DRC hotspots on CircuitNet at 1.75% NRMSE and 3.5% FPR, outperforming U-Net baselines and cutting false positives by about half.

desk verdict Table II's F1 numbers are arithmetically inconsistent with the reported TPR and Precision, so MAGNet's claimed superiority is not established despite a coherent architecture. read the letter →

arxiv 2506.07126 v1 pith:7PJHD7VI submitted 2025-06-08 cs.AR cs.AI

Weihan Lu , Hong Cai Chen This is my paper

classification cs.ARcs.AI

keywords designrulecheckingDRChotspotpredictionU-Netgraphneuralnetworkmulti-scaleconvolutionattentionmechanismCircuitNetVLSIlayoutverification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Design rule checking (DRC) is the verification step that ensures a chip layout can be manufactured, and it is expensive to run on full-chip designs. This paper claims that MAGNet, a hybrid network merging a multi-scale, attention-enhanced U-Net with a graph neural network over tile-local pin graphs, predicts DRC violation hotspots on the CircuitNet dataset at 1.75% NRMSE, 98.11% TPR, and 3.5% FPR. The authors argue this beats U-Net-based baselines and cuts false positives by roughly half while losing less than one point of recall. The mechanism is a graph-to-grid map that aligns pin-connectivity features with layout image features and lets the structural branch guide the U-Net's attention. A label-amplification pretraining stage is credited with making the model sensitive to very sparse violation patterns.

What carries the argument

MAGNet's carrying mechanism is a map-guided fusion of two feature streams. The image stream, MD-Unet, replaces standard convolutions with Multi-Scale Convolution Modules that run $3\times3$, $5\times5$, and $7\times7$ kernels in parallel and combine them by weighted sum, and applies dynamic channel and spatial attention to the final encoder feature map. The structural stream builds a pixel-aligned tile graph: each $256\times256$ layout tile becomes a graph whose nodes are pins carrying coordinate, layer, and obstacle-density features, edges connect same-layer pins within 1 $\mu$m with distance and relative-direction attributes, and three graph-convolution layers pass messages along those edges. A graph-to-grid map projects the GNN output back to the image grid and feeds into the attention modules, so topological cues directly modulate spatial feature weighting before a discriminator head fuses the two $256\times256$ output maps.

What would settle it

Take the reported discrete metrics for MAGNet at face value: with TPR (recall) = 98.11% and Precision = 91.00%, the F1 score is $2 \times 0.9811 \times 0.9100 / (0.9811 + 0.9100) \approx 0.944$, not the reported 69.72%. Recomputing F1 from the raw confusion matrix, or rerunning the evaluation with the train/test split stated, would settle whether the claimed false-positive reduction is real.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that spatial and structural layout information are complementary for DRC hotspot prediction and can be combined with spatial consistency: MD-Unet extracts pixel-level predictions, the tile-based GNN extracts local pin-topology embeddings, and a shared graph-to-grid map aligns the two streams for fusion. The paper reports that this fusion produces the lowest average normalized root mean square error (NRMSE, 1.75%) among compared models, a 98.11% TPR, and a 3.5% FPR that is roughly half of MD-Unet's 10.34%, with the highest F1 and precision in the discrete comparison. In the paper's telling, the GNN branch succeeds by guiding the U-Net's channel and spatial attention toward electrically and structurally critical areas, not simply by adding a second classifier.

Load-bearing premise

All comparative claims rest on the assumption that every number in Table II comes from a single correct evaluation on a held-out test split of CircuitNet that was never used in training, with TPR, precision, and F1 computed from the same confusion matrix.

Editorial extensions

If this is right

If the reported numbers hold, MAGNet can act as a fast pre-screen in the design flow, flagging DRC hotspots at 98% recall so that expensive rule-based verification is run only on a small candidate set.
The roughly 50% FPR reduction with under 1% TPR loss indicates that structural pin-connectivity cues can suppress false alarms without sacrificing hotspot recall, which is the operating point verification teams want.
The two-stage training recipe, label amplification for MD-Unet pretraining followed by frozen-GNN adaptation and then joint fine-tuning, gives a practical schedule for learning from extremely sparse violation labels.
Because the graph is built per tile and processed independently, the fusion approach remains parallelizable and scalable to full-chip layouts rather than requiring one global layout graph.
The NRMSE drop from 2.86% (MD-Unet) to 1.75% (MAGNet) implies the graph branch adds distribution-level information about violation density that image-only features do not carry.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, the same tile-graph-plus-image fusion recipe could transfer to other sparse physical-design prediction tasks, such as pin-access prediction or congestion maps, where small isolated hotspots matter more than global image fidelity.
Beyond the paper, if the F1 inconsistency is resolved and the comparison rerun, the claimed advantage may rest more on the FPR/F1 trade-off than on the NRMSE gain, so the graph branch's contribution should be evaluated separately for each metric.
Beyond the paper, applying graph-guided attention at multiple decoder levels rather than only the final encoder layer is a testable extension that may recover the small, fragmented hotspot areas the authors acknowledge are missed.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague, quick read on MAGNet. The architecture is plausible and the paper is clearly written, but the quantitative evidence for the headline claim is internally broken. The F1 values in Table II cannot be derived from the accompanying TPR and Precision columns using the paper's own Eq. (11). For MAGNet, TPR=0.9811 and Precision=0.9100 imply F1≈0.944, not the reported 0.6972. The same holds for MD-Unet and RouteNet. So every row in that table has at least one wrong entry. That alone undermines the claim of significant outperformance.

What's actually new: a tile-based graph construction derived from PGNN's Model I, a GNN branch with a spatial map to align back to the image, and a map-guided attention fusion into a U-Net with multi-scale convolution and CBAM-style attention. The parts are standard (Inception, CBAM, U-Net+GNN from Baek et al.), but the integration is a reasonable, reproducible design. The authors also acknowledge limitations—complexity, irregular batching, and noise from incremental training—which is more honest than many such papers.

Soft spots beyond the F1 arithmetic: the paper never states the train/test split on CircuitNet, gives no error bars or significance tests, and the label amplification factor is described as 10 in Section IV-A and as 100 in Section IV-D. The MD-Unet vs MAGNet comparison is not a clean ablation: MD-Unet comes from stage-1 training, while MAGNet goes through stage-2 joint training, so the GNN benefit could be an artifact of extra training. Also, on SSIM and AUC, MAGNet is slightly worse than MD-Unet (97.93 vs 98.11 SSIM; 0.82 vs 0.83 AUC), which the authors don't mention when claiming all-around superiority.

This is a paper for the ML-for-EDA crowd, as an incremental method on CircuitNet. A serious referee could ask for a re-evaluation with consistent metrics, a described split, and a proper ablation. As-is, I wouldn't cite it or rely on the numbers. My recommendation: send it to peer review with a strong expectation that it will be rejected or require major revision unless the authors fix the evaluation.

Referee Report

5 major / 7 minor

Summary. The manuscript proposes MAGNet, a hybrid DRC violation detection model that combines an attention- and multi-scale-augmented U-Net (MD-Unet) with a tile-based graph neural network, using a map-guided fusion mechanism to combine pixel-level and topological features. The authors train on the CircuitNet dataset and compare against RouteNet, ibUnet, Enhanced U-Net, and J-Net, reporting continuous metrics (NRMSE, SSIM) and discrete metrics (FPR, TPR, Accuracy, AUC, F1, Precision). The central claim is that MAGNet significantly outperforms all baselines by integrating spatial, semantic, and structural information, with a particularly strong reduction in false positive rate compared with MD-Unet.

Significance. If the reported results were reliable, the proposed architecture would be a useful addition to ML-based DRC because the tile-based graph construction and map-guided attention offer a plausible mechanism for injecting connectivity information into image-based prediction, and the paper has the merit of evaluating on a public benchmark with several metrics. However, the quantitative evidence as presented is internally inconsistent, the training protocol is contradictory, and the experimental design is under-specified. As a result, the central claim of significant improvement over existing models is not currently established; the contribution is at this stage an architectural proposal without verified empirical support.

major comments (5)

[§IV-C, Table II and Eq. (11)] The rows of Table II are arithmetically incompatible with Eq. (11), which defines F1 = 2*Precision*Recall/(Precision+Recall) with Recall=TPR. For MAGNet, TPR=98.11% and Precision=91.00% imply F1≈94.4%, not the reported 69.72. For MD-Unet, TPR=99.53% and Precision=81.66% imply F1≈89.7%, not 64.23. For RouteNet, TPR=82.1% and Precision=59.16% imply F1≈68.8%, not 55.70. Unless the metrics are macro-averaged per image rather than pooled, and this is nowhere stated, at least one entry in each row is wrong. Because the paper's headline conclusion that MAGNet significantly outperforms the baselines rests on Tables I and II, this internal inconsistency directly undermines the central empirical claim.
[§IV-A vs §IV-D] The label amplification factor is stated as 10 in the Stage-1 training description ('all label values are amplified by a factor of 10') but as 100 in the experimental analysis ('the multiplication of DRC labels by a factor of 100'). This is not a purely cosmetic discrepancy: the amplification factor is invoked in §IV-D as the explanation for MD-Unet's very high TPR. The paper must state a single value and specify exactly when it is applied; as written, the contradiction prevents reproduction of the training schedule that is central to the claimed sensitivity to sparse violation patterns.
[§IV-B and §IV-C] No train/validation/test split is described for the CircuitNet experiments, and no error bars, confidence intervals, or significance tests accompany any reported metric. The text says only that the evaluation is conducted on the CircuitNet dataset, without stating how many layouts were used, how tiles were partitioned, whether all methods were evaluated on the same split, or how the 0.1 binarization threshold was chosen. Since some reported gaps between methods are small (e.g., SSIM 97.72 vs 98.11, Accuracy 99.22 vs 98.89), the claimed superiority cannot be assessed from the point estimates as reported.
[Table II, AUC column] The AUC values in Table II are not discussed in the text, and they do not support the claim of overall superiority: MAGNet's AUC is 0.82, lower than RouteNet's 0.93, J-Net's 0.84, and MD-Unet's 0.83. If these numbers are correct, a threshold-independent ranking metric places MAGNet below several baselines, which is difficult to reconcile with the statement that MAGNet 'significantly outperforms' them. The paper should either explain this discrepancy or remove the AUC column from the evidence for the central claim.
[§III-A and §IV-C] The contribution of the GNN branch is not isolated by a controlled ablation. The comparison of MAGNet with MD-Unet in Table II is presented as evidence that graph-guided attention reduces FPR, but the two models are trained under different protocols (single-stage training for MD-Unet vs two-stage joint training for MAGNet), and no experiment holds the backbone, training schedule, and threshold fixed while adding only the graph branch. The observed differences could therefore be due to additional model capacity or to the different training strategy rather than to graph fusion.

minor comments (7)

[§IV-B] 'True Positive Rate (TPF)' should read 'True Positive Rate (TPR)', and 'RoutNet' in Table II should be 'RouteNet'.
[§III-C and §IV-B] Equation (9) is used twice: once for map-guided spatial attention in §III-C and once for NRMSE in §IV-B. Please renumber the equations consistently.
[Fig. 11] The training-loss plot lacks axis labels and a clear description of what is being compared; without this information, the claim that MAGNet's loss is 'approximately one order of magnitude less than MD-Unet alone' is difficult to verify.
[Abstract and throughout] The model name is written inconsistently as both 'MAGNet' and 'MAGnet'; please standardize to a single spelling.
[§IV-C] The sentence 'the change in continuous indicators (such as accuracy, precision, and F1-score)' is inaccurate because Accuracy, Precision, and F1 are discrete indicators from Table II, not continuous indicators from Table I.
[§III-D] The final prediction layers are called a 'discriminator head,' but no adversarial or discriminator training objective is described; a less suggestive term such as 'prediction head' would avoid confusion.
[§IV-C] The text states that 'DRC is to identify hotspots, and hotspots are discrete variables' after also treating the DRC violation count as a continuous variable; the relationship between the continuous density prediction and the binary hotspot prediction needs a clearer explanation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MAGNet's DRC predictions are trained and evaluated on the external CircuitNet benchmark, and the reported weaknesses are correctness risks rather than circular derivation.

full rationale

MAGNet's claimed contribution is a hybrid U-Net/GNN architecture trained on CircuitNet layout features to predict DRC hotspot maps. None of the prediction results is defined in terms of the headline metrics, and no parameter is fitted to the reported test statistics in a way that makes a later claim a tautology. The 0.1 threshold is described as a fixed operating point 'chosen to ensure high recall' (Section III-D), but the resulting TPR/FPR still depend on the model's output scores and are not arithmetically forced by the threshold alone, so this is an evaluation-protocol concern rather than a circular step. Reference [7], possibly sharing an author with this paper, is used only as an external baseline and is not load-bearing for the central claim. The paper's serious problems are internal inconsistencies and under-reporting: Table II's F1 values are incompatible with the stated TPR and Precision under Eq. (11), the train/test split is never described, error bars/significance tests are absent, and the label amplification factor is stated as 10 in Section IV-A but 100 in Section IV-D. These are correctness and reproducibility risks, not circularity under the enumerated patterns, and they do not make the derivation equivalent to its inputs by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the CircuitNet ground truth, the PGNN connectivity model, and the stated but incomplete training details. The model itself contains thousands of trained weights, which are not enumerated here; the free parameters listed are the hand-chosen control knobs that directly shape the reported metrics.

free parameters (2)

Binarization threshold = 0.1
Hand-chosen threshold applied to sigmoid output to favor recall; directly determines TPR, FPR, F1, Precision in Table II.
Label amplification factor = 10 (stated) / 100 (stated in analysis)
Stage-1 training multiplies sparse DRC labels to boost gradient flow; the paper gives two different values in different sections, a direct inconsistency.

assumptions (4)

domain assumption CircuitNet provides accurate, representative ground-truth DRC violation labels.
All performance numbers are computed against CircuitNet labels; no independent verification of labels is given (Section IV-C).
domain assumption A two-layer orthogonal wiring model (one vertical, one horizontal) is sufficient to represent pin connectivity.
Graph construction assumes multi-layer bidirectional wiring with first layer vertical, second horizontal (Section III-B).
domain assumption PGNN Model I correctly characterizes the four pin-connectivity factors (shape, connection mode, mutual influence, pin count).
The pixel-based graph adopts Model I from PGNN as the feature model (Section III-B).
ad hoc to paper U-Net and GNN features are complementary, and fusing them improves DRC prediction quality.
This is the core design premise of MAGNet; it is asserted and tested only through the final performance comparison, without isolating the fusion's contribution in an ablation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MAGNet: A Multi-Scale Attention-Guided Graph Fusion Network for DRC Violation Detection." pith.science (2026). https://pith.science/paper/7PJHD7VI

@misc{pith2026250607126,
  author       = {Pith},
  title        = {Pith review of: MAGNet: A Multi-Scale Attention-Guided Graph Fusion Network for DRC Violation Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7PJHD7VI}},
  note         = {Machine review of arXiv:2506.07126}
}

read the original abstract

Design rule checking (DRC) is of great significance for cost reduction and design efficiency improvement in integrated circuit (IC) designs. Machine-learning-based DRC has become an important approach in computer-aided design (CAD). In this paper, we propose MAGNet, a hybrid deep learning model that integrates an improved U-Net with a graph neural network for DRC violation prediction. The U-Net backbone is enhanced with a Dynamic Attention Module (DAM) and a Multi-Scale Convolution Module (MSCM) to strengthen its capability in extracting fine-grained and multi-scale spatial features. In parallel, we construct a pixel-aligned graph structure based on chip layout tiles, and apply a specialized GNN to model the topological relationships among pins. During graph construction, a graph-to-grid mapping is generated to align GNN features with the layout image. In addition, a label amplification strategy is adopted during training to enhance the model's sensitivity to sparse violation patterns. Overall, MAGNet effectively combines spatial, semantic, and structural information, achieving improved prediction accuracy and reduced false positive rates in DRC hotspot detection. Subsequently, through incremental training, we achieve a more sensitive discrimination ability for hotspots. The results demonstrate that, in comparison with ibUnet, RouteNet, and J-Net, MAGnet significantly outperforms these models, achieving substantial improvements in overall performance.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 28 canonical work pages

[1]

Recent progress of integrated circuits and optoelectronic chips,

Y. Hao, S. Xiang, G. Han, et al., "Recent progress of integrated circuits and optoelectronic chips," Science China Information Sciences, vol. 64, no. 10, pp. 201401, 2021

work page 2021
[2]

Machine Learning for Electronic Design Automation: A Survey,

G. Y. Huang, J. Hu, Y. He, et al., "Machine Learning for Electronic Design Automation: A Survey," ACM Transactions on Design Automation of Electronic Systems, vol. 26, no. 5, pp. 1-46, 2021

work page 2021
[3]

- RouteNet: Routability prediction for Mixed-Size Designs Using Convolutional Neural Network,

Z. Xie, Y. H. Huang, G. Q. Fang, et al, "- RouteNet: Routability prediction for Mixed-Size Designs Using Convolutional Neural Network," in Proc. Int. Conf. Computer -Aided Design (ICCAD), San Francisco, CA, USA, 2018, pp. -

work page 2018
[4]

Design Rule Checking with a CNN Based Feature Extractor,

L. Francisco, J. Doe, R. Smith, et al., "Design Rule Checking with a CNN Based Feature Extractor," in Proc. 2020 ACM/IEEE Workshop on Machine Learning for CAD (ML4CAD), San Diego, CA, USA, 2020, pp. -

work page 2020
[5]

DRC Violation Prediction After Global Route Through Convolutional Neural Network,

W.-T. Hung, Y. -G. Chen, J. -G. Lin, et al., "DRC Violation Prediction After Global Route Through Convolutional Neural Network," IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 31, no. 9, pp. 1425–1438, 2023

work page 2023
[6]

Pin Accessibility and Routing Congestion Aware DRC Hotspot Prediction Using Graph Neural Network and U -Net,

K. Baek, H. Park, S. Kim, et al., "Pin Accessibility and Routing Congestion Aware DRC Hotspot Prediction Using Graph Neural Network and U -Net," in Proc. 41st IEEE/ACM Int. Conf. Computer -Aided Design (ICCAD), Austin, TX, USA, 2022, pp. -

work page 2022
[7]

A Standard Deviation Incorporated Attention Module Enhanced U-Net for Routability Prediction,

Y. Chen, D. Zhang, and H. Chen, "A Standard Deviation Incorporated Attention Module Enhanced U-Net for Routability Prediction," in 9th Int. Conf. Electronic Technology and Information Science (ICETIS), Beijing, China, 2024, pp. -

work page 2024
[8]

A Lightweight Inception Boosted U-Net Neural Network for Routability Prediction,

H. Li, Y. Huo, Y. Wang, et al., "A Lightweight Inception Boosted U-Net Neural Network for Routability Prediction," in 2nd Int. Symp. Electronics Design Automation (ISEDA), Taipei, Taiwan, 2024, pp. -

work page 2024

Show all 32 references

[9]

DRC Hotspot Prediction at Sub - 10nm Process Nodes Using Customized Convolutional Network,

R. Liang, H. Xiang, D. Pandey, et al., "DRC Hotspot Prediction at Sub - 10nm Process Nodes Using Customized Convolutional Network," in Proc. 2020 Int. Symp. Physical Design (ISPD), San Francisco, CA, USA, 2020, pp. -

2020
[10]

Circuit as Set of Points,

J. Zou, X. G. Wang, J. H. Guo, et al., "Circuit as Set of Points," in 37th Conf. Neural Information Processing Systems (NeurIPS), New Orleans, LA, USA, 2023, pp. -

2023
[11]

LHNN: Lattice Hypergraph Neural Network for VLSI Congestion Prediction,

B. W. Wang, L. Zhang, M. Li, et al., "LHNN: Lattice Hypergraph Neural Network for VLSI Congestion Prediction," in 59th ACM/IEEE Design Automation Conf. (DAC), San Francisco, CA, USA, 2022, pp. 1297–1302

2022
[12]

Early Stage DRC Prediction Using Ensemble Machine Learning Algorithms,

R. Islam, "Early Stage DRC Prediction Using Ensemble Machine Learning Algorithms," IEEE Canadian Journal of Electrical and Computer Engineering, vol. 45, no. 4, pp. 354–364, 2022

2022
[13]

Explainable DRC Hotspot Prediction with Random Forest and SHAP Tree Explainer,

W. Zeng, A. Davoodi, and R. O. Topaloglu, "Explainable DRC Hotspot Prediction with Random Forest and SHAP Tree Explainer," in Design, Automation and Test in Europe Conf. and Exhibition (DATE), Grenoble, France, 2020, pp. 1151–1156

2020
[14]

CircuitNet: An Open-Source Dataset for Machine Learning in VLSI CAD Applications With Improved Domain - Specific Evaluation Metric and Learning Strategies,

Z. Chai, Y. Zhao, W. Liu, et al., "CircuitNet: An Open-Source Dataset for Machine Learning in VLSI CAD Applications With Improved Domain - Specific Evaluation Metric and Learning Strategies," IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , vol. ...

2023
[15]

U -Net: Convolutional Networks for Biomedical Image Segmentation,

O. Ronneberger, P. Fischer, and T. Brox, "U -Net: Convolutional Networks for Biomedical Image Segmentation," in Medical Image Computing and Computer -Assisted Intervention (MICCAI), Munich, Germany, 2015, pp. 234–241

2015
[16]

A General Survey on Attention Mechanisms in Deep Learning,

G. Brauwers and F. Frasincar, "A General Survey on Attention Mechanisms in Deep Learning," IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 4, pp. 3279–3298, 2023

2023
[17]

CBAM: Convolutional Block Attention Module,

S. Woo, J. Park, J. -Y. Lee, and I. S. Kweon, "CBAM: Convolutional Block Attention Module," in European Conf. Computer Vision (ECCV), Munich, Germany, 2018, pp. 3–19

2018
[18]

Competitive Multi -scale Convolution,

Z. Liao and G. Carneiro, "Competitive Multi -scale Convolution," arXiv preprint, 2015, arXiv:1511.05635

2015 arXiv
[19]

Going Deeper with Convolutions,

C. Szegedy, W. Liu, Y. Jia, et al., "Going Deeper with Convolutions," in IEEE Conf. Computer Vision and Pattern Recognition (CVPR), Boston, MA, USA, 2015, pp. 1–9

2015
[20]

FLUTE: Fast Lookup Table Based Wirelength Estimation Technique,

C. Chu, "FLUTE: Fast Lookup Table Based Wirelength Estimation Technique," in IEEE/ACM Int. Conf. Computer Aided Design (ICCAD), San Diego, CA, USA, 2004, pp. 696–701

2004
[21]

FastRoute: A Step to Integrate Global Routing into Placement,

M. Pan and C. Chu, "FastRoute: A Step to Integrate Global Routing into Placement," in 2006 IEEE/ACM Int. Conf. Computer Aided Design (ICCAD), San Francisco, CA, USA, 2006, pp. 464–471

2006
[22]

The Graph Neural Network Model,

F. Scarselli, M. Gori, A. C. Tsoi, et al., "The Graph Neural Network Model," IEEE Transactions on Neural Networks, vol. 20, no. 1, pp. 61 – 80, 2009

2009
[23]

Methodology of Resolving Design Rule Checking Violations Coupled with Fully Compatible Prediction Model,

S. Kim, H. Park, K. Baek, et al., "Methodology of Resolving Design Rule Checking Violations Coupled with Fully Compatible Prediction Model," in Proc. 2024 Int. Symp. Physical Design (ISPD), Taipei, Taiwan, 2024, pp. –

2024
[24]

Novel Transformer Model Based Clustering Method for Standard Cell Design Automation,

C.-T. Ho, A. Chandna, D. Guan, et al., "Novel Transformer Model Based Clustering Method for Standard Cell Design Automation," in Proc. 2024 Int. Symp. Physical Design (ISPD), Taipei, Taiwan, 2024, pp. -

2024
[25]

Graph Attention Networks,

P. Veličković, G. Cucurull, A. Casanova, et al., "Graph Attention Networks," arXiv preprint, 2017, arXiv:1710.10903

2017 arXiv
[26]

Dynamic Graph CNN for Learning on Point Clouds,

Y. Wang, Y. Sun, Z. Liu, et al., "Dynamic Graph CNN for Learning on Point Clouds," arXiv preprint, 2018, arXiv:1801.07829

2018 arXiv
[27]

UNet++: A Nested U-Net Architecture for Medical Image Segmentation,

Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, et al., "UNet++: A Nested U-Net Architecture for Medical Image Segmentation," arXiv preprint, 2018, arXiv:1807.10165

2018 arXiv
[28]

Attention U-Net: Learning Where to Look for the Pancreas,

O. Oktay, J. Schlemper, L. L. Folgner, et al., "Attention U-Net: Learning Where to Look for the Pancreas," arXiv preprint, 2018, arXiv:1804.03999

2018 arXiv
[29]

Document -level relation extraction with multi -layer heterogeneous graph attention network,

Nianbin Wang, Tiantian Chen, Chaoqi Ren, et al., "Document -level relation extraction with multi -layer heterogeneous graph attention network," Engineering Applications of Artificial Intelligence , vol. 123, Part A, pp. 106212, 2023

2023
[30]

Inductive Representation Learning on Large Graphs,

W. L. Hamilton, R. Ying, and J. Leskovec, "Inductive Representation Learning on Large Graphs," in Proc. 31st Int. Conf. Neural Information Processing Systems (NeurIPS), Long Beach, CA, USA, 2017, pp. -

2017
[31]

DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,

L. C. Chen, G. Papandreou, I. Kokkinos, et al., "DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 40, no. 4, pp. 834–848, 2018

2018
[32]

GRANNITE: Graph Neural Network Inference for Transferable Power Estimation,

Y. Zhang, H. Ren, and B. Khailany, "GRANNITE: Graph Neural Network Inference for Transferable Power Estimation," in 57th ACM/IEEE Design Automation Conf. (DAC), San Francisco, CA, USA, 2020, pp. 1–6. WeiHan Lu was born in China in 2004. He is currently an undergraduate student...

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Recent progress of integrated circuits and optoelectronic chips,

Y. Hao, S. Xiang, G. Han, et al., "Recent progress of integrated circuits and optoelectronic chips," Science China Information Sciences, vol. 64, no. 10, pp. 201401, 2021

work page 2021

[2] [2]

Machine Learning for Electronic Design Automation: A Survey,

G. Y. Huang, J. Hu, Y. He, et al., "Machine Learning for Electronic Design Automation: A Survey," ACM Transactions on Design Automation of Electronic Systems, vol. 26, no. 5, pp. 1-46, 2021

work page 2021

[3] [3]

- RouteNet: Routability prediction for Mixed-Size Designs Using Convolutional Neural Network,

Z. Xie, Y. H. Huang, G. Q. Fang, et al, "- RouteNet: Routability prediction for Mixed-Size Designs Using Convolutional Neural Network," in Proc. Int. Conf. Computer -Aided Design (ICCAD), San Francisco, CA, USA, 2018, pp. -

work page 2018

[4] [4]

Design Rule Checking with a CNN Based Feature Extractor,

L. Francisco, J. Doe, R. Smith, et al., "Design Rule Checking with a CNN Based Feature Extractor," in Proc. 2020 ACM/IEEE Workshop on Machine Learning for CAD (ML4CAD), San Diego, CA, USA, 2020, pp. -

work page 2020

[5] [5]

DRC Violation Prediction After Global Route Through Convolutional Neural Network,

W.-T. Hung, Y. -G. Chen, J. -G. Lin, et al., "DRC Violation Prediction After Global Route Through Convolutional Neural Network," IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 31, no. 9, pp. 1425–1438, 2023

work page 2023

[6] [6]

Pin Accessibility and Routing Congestion Aware DRC Hotspot Prediction Using Graph Neural Network and U -Net,

K. Baek, H. Park, S. Kim, et al., "Pin Accessibility and Routing Congestion Aware DRC Hotspot Prediction Using Graph Neural Network and U -Net," in Proc. 41st IEEE/ACM Int. Conf. Computer -Aided Design (ICCAD), Austin, TX, USA, 2022, pp. -

work page 2022

[7] [7]

A Standard Deviation Incorporated Attention Module Enhanced U-Net for Routability Prediction,

Y. Chen, D. Zhang, and H. Chen, "A Standard Deviation Incorporated Attention Module Enhanced U-Net for Routability Prediction," in 9th Int. Conf. Electronic Technology and Information Science (ICETIS), Beijing, China, 2024, pp. -

work page 2024

[8] [8]

A Lightweight Inception Boosted U-Net Neural Network for Routability Prediction,

H. Li, Y. Huo, Y. Wang, et al., "A Lightweight Inception Boosted U-Net Neural Network for Routability Prediction," in 2nd Int. Symp. Electronics Design Automation (ISEDA), Taipei, Taiwan, 2024, pp. -

work page 2024