Pith. sign in

REVIEW 3 major objections 4 minor 39 references

A spatio-temporal graph network predicts multi-cycle fault impact probabilities in sequential circuits from five cycles of prior features, cutting simulation time by over 10x with 5-cycle MAE down to 0.024.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-04 23:51 UTC pith:KYNO2VCT

load-bearing objection Plausible ST-GCN approach to fast FIP prediction, but the headline numbers are not clearly held-out and the abstract pairs speed from one configuration with accuracy from another; the TPI validation is circular. the 3 major comments →

arxiv 2509.06289 v1 pith:KYNO2VCT submitted 2025-09-08 cs.LG cs.ARcs.ET

A Spatio-Temporal Graph Neural Networks Approach for Predicting Silent Data Corruption inducing Circuit-Level Faults

Shaoqi Wei , Senling Wang , Hiroshi Kai , Yoshinobu Higami , Ruijun Ma , Tianming Ni , Xiaoqing Wen , Hiroshi Takahashi This is my paper
classification cs.LG cs.ARcs.ET
keywords silent data errorsfault impact probabilityspatio-temporal graph convolutional networksequential circuit testingtest point insertionfunctional safetytestability metricsEDA integration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Silent data errors are hard to catch because the underlying defects often only show up after many clock cycles of functional operation, and full fault simulation for every candidate fault is too slow. The paper proposes a spatio-temporal graph convolutional network that treats a gate-level circuit as a graph whose edges carry cycle-by-cycle signal features, then predicts the fault impact probability (FIP) for the next five or ten cycles from the previous five. On standard benchmark circuits it reports mean absolute error as low as 0.024 for five-cycle predictions while cutting simulation time by more than an order of magnitude. The same model accepts either cheap testability metrics or expensive fault-simulation data as input, and a test-point-selection case study uses predicted FIP to place observation points that detect long-cycle faults earlier. If the transfer to larger designs holds, this gives test engineers a quantitative risk score for hard-to-detect faults without running exhaustive multi-cycle simulation.

Core claim

The central claim is that long-cycle Fault Impact Probability — the fraction of test patterns that carry a given fault's effect to an observation point within a given clock cycle — is learnable from recent circuit history. The model maps five cycles of per-signal-line features (either testability metrics or simulated FIP values) plus the gate-level topology to FIP values five or ten cycles later. The reported accuracy (MAE 0.024 for 5-cycle predictions with simulation-based features; around 0.033 average with metric-based features) is close enough to the simulated ground truth that the authors use predicted FIP to rank cycle-sensitive faults and select observation points.

What carries the argument

The central object is the ST-Graph: a gate-level netlist turned into a time-evolving graph where nodes are gates and flip-flops, edges are signal lines carrying cycle-by-cycle feature vectors, and a fixed time window of edge features is fed through a residual gated graph convolutional encoder (spatial) and a graph-transformer multi-head attention encoder (temporal), whose outputs are summed and decoded into per-node FIP vectors. The gating coefficient suppresses redundant paths while the attention weight highlights globally relevant neighbors, together capturing the multi-cycle propagation that static single-cycle GCNs miss.

Load-bearing premise

The previous five cycles of testability or simulated fault-impact data contain enough information to predict the next five to ten cycles, and the mapping learned on small benchmark circuits still holds when observation points are moved—a step the test-point study does not verify with fresh fault simulation.

What would settle it

Run full fault simulation on the same circuits after selecting observation points with the model and compare the actual FIP change to the predicted change; a systematic gap (e.g., average MAE above 0.05 on circuits not in the training split) would falsify the transfer claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A trained ST-GCN can rank faults by their long-cycle detection probability without running the full multi-cycle simulation, so test effort can be concentrated where risk is highest.
  • Using testability metrics as input cuts ST-Graph construction time from roughly exponential to linear growth in circuit size while keeping average MAE around 0.033 for 5-cycle predictions, an explicit efficiency-accuracy trade-off.
  • Test point insertion guided by predicted FIP raises the average FIP within the first four clock cycles on the studied circuits, meaning cycle-sensitive hard-to-detect faults become observable earlier in functional testing.
  • Because the observation point set is a parameter of the FIP definition, the same framework covers functional testing, multi-cycle logic BIST, and test point insertion without retraining on the test architecture.
  • GPU inference of about 4 seconds per 10,000 gates, with little added cost for extending the horizon from 5 to 10 cycles, makes the model usable inside EDA flows that currently rely on slower structural analysis.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same ST-Graph encoding could be applied to system-level reliability if node features were extended to functional blocks and primary-output weighting, moving the current gate-level FIP toward the system-level SDE risk the authors list as future work.
  • The reported 10x+ speedup is measured on small benchmark circuits; the practical gain could grow on industrial designs, where fault simulation cost scales superlinearly with circuit size and pattern count, or shrink if the learned mapping does not transfer.
  • A direct test of the TPI claim would be to run fault simulation after the chosen observation points are inserted; the paper currently validates only predicted FIP changes, not actual detection improvement.
  • Cycle-sensitive faults identified by the model could also guide functional test length selection, not just physical observation-point placement, because their FIP rises slowly over cycles.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a spatio-temporal graph convolutional network (ST-GCN) for predicting per-cycle Fault Impact Probability (FIP) on signal lines of sequential circuits. Gate-level netlists are converted into ST-Graphs with gate-type node features and time-varying edge features derived either from testability metrics (SCOAP/COP) or from multi-cycle fault simulation. A spatial gated-GCN encoder and a temporal graph-Transformer encoder are combined, and a decoder predicts FIP for the next 5 or 10 cycles. Experiments on 18 ISCAS-89 circuits report RMSE/MAE under uniform and sparse training-set sampling, an ablation study over the main architectural components, and runtime comparisons for graph conversion, training, and inference. A test-point insertion case study uses the TM-10-U model to greedily select DFF observation points based on predicted reductions in cycle-sensitive faults.

Significance. If the reported accuracy is genuinely held-out, the ST-GCN approach would provide a fast surrogate for multi-cycle fault simulation, with clear application to functional test and test-point insertion. The paper's strengths include the explicit spatio-temporal formulation of sequential-circuit fault propagation, the support for two feature-modeling strategies with an efficiency-accuracy trade-off, and an ablation study that isolates the contributions of time encoding, the spatial encoder, and the temporal encoder. The empirical claims are, however, not yet fully supported: the train/test split is not mapped onto the reported per-circuit tables, and the headline numbers pair accuracy from one feature pipeline with speedup from another. These issues must be resolved before the central claim can be accepted.

major comments (3)
  1. [§5.1, Tables 1 and 2] The dataset description in §5.1 states that circuits are sorted by size and that uniform/sparse sampling selects every other/every two circuits for training. Tables 1 and 2, however, list per-circuit RMSE/MAE for all 18 circuits with a single Average row and no marker indicating which circuits were held out. If those averages include training circuits, the reported MAE of 0.024 (Table 1, FT-5-U average) is not a valid estimate of generalization and cannot support the claim that the model transfers to larger designs. Please mark train/test membership in the tables, report separate test-only averages, and clarify whether any per-circuit row is a held-out circuit.
  2. [Abstract, §5.3.1, Tables 1 and 2] The abstract states that the method 'reduces simulation time by more than 10x while maintaining high accuracy (mean absolute error 0.024 for 5-cycle predictions).' The 10x speedup in Fig. 7.1 is for ST-Graph conversion using testability metrics, whereas the 0.024 MAE is the simulation-FIP-based model FT-5-U in Table 1. The testability-metric models in Table 2 have average MAE of 0.0333 for TM-5-U, so no single configuration simultaneously achieves the quoted speedup and accuracy. The abstract and conclusion should be reworded to attribute the speedup to the testability-metric pipeline and the accuracy to the simulation-FIP pipeline, or a single configuration should be reported that achieves both.
  3. [§6, Fig. 8.2, Fig. 9] The test-point insertion case study validates the method using the model's own predicted FIP as the outcome measure. The greedy procedure in Fig. 8.2 repeatedly applies the TM-10-U model to predict FIP before and after temporarily designating DFFs as observation points, and Fig. 9 reports average FIP over cycles as evidence of improvement. Because no fault simulation is run after adding observation points, the claimed improvement in detecting long-cycle, hard-to-detect faults is not independently verified. Moreover, the TM model was trained with observation points limited to primary outputs (as stated in §5.1), so its predictions under the modified observation set are an extrapolation. Please either run ground-truth fault simulation for at least a subset of the selected/DUT circuits, or explicitly rephrase the case study as a demonstration of predicted-FIP-based selection rather than ver
minor comments (4)
  1. [§5.4, Table 3] The table reporting the ablation results is introduced as 'Table reports' without a number; it should be numbered (e.g., Table 3) and referenced consistently. Also, only average RMSE/MAE are shown; reporting per-circuit results or variance would help assess stability.
  2. [Tables 1-2 and §5.1] Model names are inconsistent: Table 1 uses 'FT-5-U' while §5.1 defines 'FIP-5-U'; Table 2 uses 'TM-5-S' but the text discusses 'MT-5-S'. Please unify the nomenclature.
  3. [§5.3.2] The sentence 'We observe from the results of the MT-5-S and MT-10-S models...' appears to be based on a comparison of RMSE and MAE trends but the explanation is vague. Clarify which circuits show outlier behavior and quantify the claim.
  4. [Fig. 9] It is unclear whether the reported average FIP curves are computed from predicted FIP only, whether the random baseline uses the same prediction pipeline, and whether error bars or multiple random seeds were used. Please specify.

Circularity Check

1 steps flagged

Core FIP regression is a standard supervised task, but the test-point case study validates the model with the model's own predictions, making the 'improvement' demonstration self-referential.

specific steps
  1. self definitional [Section 6, Case Study on Test Point Selection, Fig. 8.2 and Fig. 9]
    "By analyzing these trends, a class of “cycle-sensitive faults” can be identified — these faults exhibit relatively low FIP values during the initial cycles, but show a significant increase as the cycles progress. ... In the selection procedure, the algorithm iteratively evaluates each unassigned DFF by temporarily designating it as an observation point, reapplying the TM-10-U model to predict the FIP, and calculating the reduction in the number of cycle-sensitive faults. ... Figure 9 presents the average FIP across 10 clock cycles for the selected benchmark circuits. The findings indicate that"

    The selection criterion and the validation metric are the same model output. The greedy algorithm chooses DFFs that maximize the reduction in the ST-GCN's predicted FIP, then Fig. 9 reports the ST-GCN's predicted average FIP after insertion. No ground-truth fault simulation with the inserted observation points is run, so the 'improvement' is a self-consistency check: any model that can be optimized will show an increase on the very quantity it was optimized to increase. The conclusion that detection of hard-to-detect faults improves is therefore not independently established; it is definitionally tied to the model's own predictions.

full rationale

The core FIP forecasting task (Section 3.3, Eq. 2-3) is a standard supervised regression: the model is trained on input features (testability metrics or simulated FIP from previous 5 cycles) and ground-truth FIP labels from fault simulation, then evaluated by RMSE/MAE against simulated FIP. That chain is not circular. However, the test-point selection case study (Section 6) validates the method with the method itself: the greedy algorithm optimizes the reduction in the ST-GCN's predicted FIP, and the reported improvement (Fig. 9) is the ST-GCN's predicted average FIP; no ground-truth fault simulation is run after insertion. Thus the demonstration of improved detection reduces by construction to the model's own outputs. The paper also leaves the train/test mapping for Tables 1-2 unspecified: Section 5.1 says circuits are split by uniform/sparse sampling, but the tables list all 18 circuits without marking which are held out, so the 0.024 average MAE may include training circuits; this is an evaluation-validity concern rather than a demonstrated circularity. The abstract's pairing of the 10x speedup (testability-metric pipeline) with the 0.024 MAE (simulation-FIP pipeline) is a headline mismatch, not a circular step. No load-bearing self-citation is present: [39] is cited only as related work.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on a learned mapping from graph features to FIP. The network weights, window length, prediction horizon, pattern count, and test-point limit are all chosen by the authors. The FIP quantity is a redefinition of known detection probability; no new physical entity is introduced. The key unvalidated assumption is the Markov-style predictability from a 5-cycle window and the transfer to new observation points.

free parameters (5)
  • ST-GCN learnable weights W = not reported
    Network parameters optimized with MSE loss on training circuits; the prediction function is entirely learned.
  • Sliding window length m = 5
    Hand-chosen; all models use 5 cycles of input features.
  • Prediction horizon s = 5 or 10
    Hand-chosen; FIP-5 and FIP-10 variants.
  • Number of test patterns N = 10000
    Hand-chosen for generating ground-truth FIP labels via fault simulation.
  • TPI observation point limit = 2% of DFFs
    Hand-chosen for the test point selection case study.
axioms (5)
  • domain assumption FIP as defined in Eq. 1 is an appropriate proxy for fault detection risk
    The paper equates FIP with likelihood of detection and impact, but system-level impact of SDE is not modeled.
  • ad hoc to paper The last 5 cycles of FIP or testability features are sufficient to predict future FIP
    The sliding-window scheme assumes a Markov property over 5 cycles.
  • domain assumption Gate-level netlist to ST-Graph conversion preserves fault propagation information
    Graph topology and 9-dimensional gate-type one-hot features are assumed to capture all relevant structural behavior.
  • domain assumption Fault simulation with random patterns yields accurate ground-truth FIP
    Sampling error from 10,000 patterns per fault is not quantified.
  • domain assumption Learned weights generalize across benchmark circuits
    Models trained on a subset of ISCAS-89 circuits are applied to held-out circuits and larger SoCs.
invented entities (1)
  • Fault Impact Probability (FIP) no independent evidence
    purpose: A scalar per fault/cycle measuring detection frequency at observation points
    FIP is defined in Eq. 1 as the average detection indicator over patterns; it is a renamed version of standard fault detection probability, not tied to any new physical mechanism or external falsifiable prediction.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Spatio-Temporal Graph Neural Networks Approach for Predicting Silent Data Corruption inducing Circuit-Level Faults." pith.science (2026). https://pith.science/paper/KYNO2VCT

@misc{pith2026250906289,
  author       = {Pith},
  title        = {Pith review of: A Spatio-Temporal Graph Neural Networks Approach for Predicting Silent Data Corruption inducing Circuit-Level Faults},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KYNO2VCT}},
  note         = {Machine review of arXiv:2509.06289}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Silent Data Errors (SDEs) from time-zero defects and aging degrade safety-critical systems. Functional testing detects SDE-related faults but is expensive to simulate. We present a unified spatio-temporal graph convolutional network (ST-GCN) for fast, accurate prediction of long-cycle fault impact probabilities (FIPs) in large sequential circuits, supporting quantitative risk assessment. Gate-level netlists are modeled as spatio-temporal graphs to capture topology and signal timing; dedicated spatial and temporal encoders predict multi-cycle FIPs efficiently. On ISCAS-89 benchmarks, the method reduces simulation time by more than 10x while maintaining high accuracy (mean absolute error 0.024 for 5-cycle predictions). The framework accepts features from testability metrics or fault simulation, allowing efficiency-accuracy trade-offs. A test-point selection study shows that choosing observation points by predicted FIPs improves detection of long-cycle, hard-to-detect faults. The approach scales to SoC-level test strategy optimization and fits downstream electronic design automation flows.

Figures

Figures reproduced from arXiv: 2509.06289 by Hiroshi Kai, Hiroshi Takahashi, Ruijun Ma, Senling Wang, Shaoqi Wei, Tianming Ni, Xiaoqing Wen, Yoshinobu Higami.

Figure 1
Figure 1. Figure 1: Architecture of the Proposed ST-GCN-Based FIP Prediction Framework 4.1.1 Testability Metrics-Based Method The temporal feature sequence for each signal line consists of controllability metrics (CC0, CC1), observability (CO), the probability of controlling a signal line to logic 1 (C1), and the probability of observing a signal line (O). These metrics are calculated using standard testability evaluation tec… view at source ↗
Figure 2
Figure 2. Figure 2: Process of Transforming a Logic Circuit into a Spatial Graph Structure for ST-Graph Construction. (1) One-hot encoding for gate types. (2) Adjacency matrix representation of circuit connectivity. (3) Node feature matrix for gates. ST raph Converter ate level netlist Spatio Temporal raph Structure [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 29 canonical work pages · 1 internal anchor

  1. [1]

    Data Center Silent Data Errors Technical Paper

    Intel 2024. Data Center Silent Data Errors Technical Paper . Retrieved July 3, 2025 from https://www.intel.com/content/www/us/en/content- details/788204/data-center-silent-data-errors-technical-paper.html

    work page 2024

  2. [2]

    Silent ata Corruptions at Scale

    Harish Dattatraya Dixit, Sneha Pendharkar, Matt Beadon, Chris Mason, Tejasvi Chakravarthy, Bharath Muthiah, and Sriram Sankar . 1. “Silent ata Corruptions at Scale.” ar iv preprint ar iv: 1 .11 . etrieved from https://arxiv.org/abs/ 1 .11

    work page

  3. [3]

    Cores that don't count

    Peter H. Hochschild, Paul Turner, Jeffrey C. Mogul, Rama Govindaraju, Parthasarathy Ranganathan, David E. Culler, and Amin Vahdat. 2021. "Cores that don't count." In Proceedings of the Workshop on Hot Topics in Operating Systems (HotOS '21), Ann Arbor, Mic higan, USA, 9–16. Association for Computing Machinery, New York, NY, USA. https://doi.org/10.1145/34...

    work page arXiv 2021

  4. [4]

    Shaobu Wang, Guangyan Zhang, Junyu Wei, Yang Wang, Jiesheng Wu, and Qingchao Luo. 2024. Understanding Silent Data Corruption in Processors for Mitigating its Effects. ACM Trans. Archit. Code Optim. 21, 4, Article 84 (December 2024), 27 pages. https://doi.o rg/10.1145/3690825

    doi:10.1145/3690825 2024

  5. [5]

    etecting silent data corruptions in the wild

    arish attatraya ixit, Laura Boyle, autham Vunnam, Sneha endharkar, Matt Beadon, and Sriram Sankar. . “ etecting silent data corruptions in the wild.” ar iv preprint ar iv: . 8989. etrieved from https://arxiv.org/abs/ . 8989

    work page

  6. [6]

    Silent Data Corruption Estimation and Mitigation Without Fault I njection

    Moona Yakhchi, Mahdi Fazeli, and Seyyed Amir Asghari. 2022. "Silent Data Corruption Estimation and Mitigation Without Fault I njection." IEEE Canadian Journal of Electrical and Computer Engineering 45, 3 (2022), 318 –327. https://doi.org/10.1109/ICJECE.2022.3189043

    work page arXiv 2022

  7. [7]

    Shaobu Wang, Guangyan Zhang, Junyu Wei, Yang Wang, Jiesheng Wu, and Qingchao Luo. 2023. Understanding Silent Data Corruptions in a Large Production CPU Population. In Proceedings of the 29th Symposium on Operating Systems Principles (SOSP '23). Association for Computing Machinery, New York, NY, USA, 216–230. https://doi.org/10.1145/3600006.3613149

    arXiv 2023

  8. [8]

    Harpocrates: Breaking the Silence of CPU Faults through Hardware -in-the-Loop Program Generation

    Nikos Karystinos, Odysseas Chatzopoulos, George-Marios Fragkoulis, George Papadimitriou, Dimitris Gizopoulos, and Sudhanva Gurumurthi. 2024. "Harpocrates: Breaking the Silence of CPU Faults through Hardware -in-the-Loop Program Generation." In Proceedings o f the 2024 ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA), 516 –531. ...

    arXiv 2024

  9. [12]

    J., Mishra, S., Marinissen , E

    Sangani, D., Kaczer, B., Weckx, P., Roussel, Ph. J., Mishra, S., Marinissen , E. J., and Gielen, G. 2024. Possible Origins, Identification, and Screening of Silent Data Corruption in Data Centers. In 2024 IEEE International Reliability Physics Symposium (IRPS), 1 -7. https://doi.org/10.1109/IRPS48228.2024.10529436

    arXiv 2024

  10. [13]

    Roussel, and Georges Gielen

    Erik Jan Marinissen, Harish Dattatraya Dixit, Shawn Blanton, Aaron Kuo, Wei Li, Subhashish Mitra, Chris Nigh, Ruben Purdy, Be n Kaczer, Dishant Sangani, Pieter Weckx, Philippe J. Roussel, and Georges Gielen. 2024. Silent Data Corruption: Test or Reliability Problem? In Proceedings of the 2024 IEEE European Test Symposium (ETS), 1–7. https://doi.org/10.110...

    work page arXiv 2024

  11. [14]

    Irith Pomeranz. 2025. Direct Search Procedure for Functional Compaction With Improved Fault Coverage. IEEE Transactions on Co mputer-Aided Design of Integrated Circuits and Systems, 44, 5 (2025), 1981 –1990. https://doi.org/10.1109/TCAD.2024.3499898

    work page arXiv 2025

  12. [15]

    Kang Zhao and Wenbo Shen. 2015. Parallel Stimulus Generation Based on Model Checking for Coherence Protocol Verification. IEE E Transactions on Very Large Scale Integration (VLSI) Systems, 23, 12 (2015), 3124 -3128. https://doi.org/10.1109/TVLSI.2014.2384040

    work page arXiv 2015

  13. [16]

    Brayton, and Giovanni De Micheli

    Siang-Yun Lee, Heinz Riener, Alan Mishchenko, Robert K. Brayton, and Giovanni De Micheli. 2022. A Simulation -Guided Paradigm for Logic Synthesis and Verification. IEEE Transactions on Computer -Aided Design of Integrated Circuits and Systems, 41, 8 (2022), 2573–2586. https://doi.org/10.1109/TCAD.2021.3108704

    work page arXiv 2022

  14. [17]

    Irith Pomeranz. 2016. Static Test Compaction for Functional Test Sequences With Restoration of Functional Switching Activity. IEEE Transactions 20 on Computer-Aided Design of Integrated Circuits and Systems, 35, 10 (2016), 1755 –1762. https://doi.org/10.1109/TCAD.2015.2512931

    work page arXiv 2016

  15. [18]

    Li, W.N., Reddy, S.M., and Sahni, S. 1988. On path selection in combinational logic circuits. Proceedings of the 25th ACM/IEE E Design Automation Conference (DAC '88), 142–147. https://doi.org/10.1109/DAC.1988.14749

    work page arXiv 1988

  16. [19]

    Irith Pomeranz. 2024. Functionally -Possible Gate -Exhaustive Bridging Faults. In 2024 IEEE International Test Conference (ITC), 31 –35. https://doi.org/10.1109/ITC51657.2024.00014

    work page arXiv 2024

  17. [20]

    Streaming Active Deep Forest for Evolving Data Stream Classification

    Yun Shao, I. Pomeranz, and S. M. Reddy. 2002. On generating high quality tests for transition faults. In Proceedings of the 11th Asian Test Symposium (ATS '02), 1–8. https://doi.org/10.1109/ATS.2002.1181676

    work page internal anchor Pith review Pith/arXiv arXiv 2002

  18. [21]

    Richard Putman and Rahul Gawde. 2006. Enhanced Timing-Based Transition Delay Testing for Small Delay Defects. In Proceedings of the 24th IEEE VLSI Test Symposium (VTS '06). IEEE Computer Society, USA, 336 –342. https://doi.org/10.1109/VTS.2006.33

    work page doi:10.1109/vts.2006.33 2006

  19. [22]

    Lin, Xijiang, Tsai, Kun -han, Wang, Chen, Kassab, Mark, Rajski, Janusz, Kobayashi, Takeo, Klingenberg, Randy, Sato, Yasuo, Hamada, Shuji, and Aikyo, Takashi. 2006. Timing -Aware ATPG for High Quality At -speed Testing of Small Delay Defects. In 2006 15th Asia n Test Symposium, 139 –

    work page 2006

  20. [23]

    Jie Jiang, Matthias Sauer, Alexander Czutro, Bernd Becker, and Ilia Polian. 2012. On the optimality of K longest path generat ion algorithm under memory constraints. In 2012 Design, Automation & Test in Europe Conference & Exhibition (DATE), 418 –423. https://doi.org/10.1109/DATE.2012.6176507

    work page arXiv 2012

  21. [24]

    Jifeng Chen and Mohammad Tehranipoor. 2013. Critical Paths Selection and Test Cost Reduction Considering Process Variations. In 2013 22nd Asian Test Symposium. 259–264. https://doi.org/10.1109/ATS.2013.55

    work page doi:10.1109/ats.2013.55 2013

  22. [25]

    M., and Kundu, S

    Seshadri, B., Pomeranz, I., Reddy, S. M., and Kundu, S. 2005. Path -oriented transition fault test generation considering operating conditions. In European Test Symposium ETS’ , –59. https://doi.org/10.1109/ETS.2005.31

    work page doi:10.1109/ets.2005.31 2005

  23. [26]

    Bhattacharya, Xiaoqing Wen, and Xijiang Lin

    Dong Xiang, Kele Shen, Bhargab B. Bhattacharya, Xiaoqing Wen, and Xijiang Lin. 2016. Thermal -Aware Small-Delay Defect Testing in Integrated Circuits for Mitigating Overkill. IEEE Transactions on Computer -Aided Design of Integrated Circuits and Systems 35, 3 (2016), 499 –512. https://doi.org/10.1109/TCAD.2015.2474365

    arXiv 2016

  24. [27]

    Murakami, A., Kajihara, S., Sasao, T., Pomeranz, I., and Reddy, S. M. 2000. Selection of potentially testable path delay faul ts for test generation. In Proceedings International Test Conference 2000 (IEEE Cat. No.00CH37159), 376 –384. https://doi.org/10.1109/TEST.2000.894227

    work page arXiv 2000

  25. [28]

    Saravanan Padmanaban and Spyros Tragoudas. 2004. A Critical Path Selection Method for Delay Testing. In Proceedings of the In ternational Test Conference on International Test Conference (ITC '04). IEEE Computer Society, USA, 232 –241

    work page 2004

  26. [29]

    Irith Pomeranz. 2020. Selecting Close -to-Functional Path Delay Faults for Test Generation. In 2020 IEEE International Test Conference (ITC), 1 –5. https://doi.org/10.1109/ITC44778.2020.9325255

    arXiv 2020

  27. [30]

    Irith Pomeranz and Yervant Zorian. 2024. Functionally Possible Path Delay Faults With High Functional Switching Activity. IEE E Trans. Very Large Scale Integr. Syst. 32, 11 (Nov. 2024), 2159–2163. https://doi.org/10.1109/TVLSI.2024.3425817

    work page arXiv 2024

  28. [31]

    Irith Pomeranz. 2012. On the detection of path delay faults by functional broadside tests. In 2012 17th IEEE European Test Sy mposium (ETS), 1–6. https://doi.org/10.1109/ETS.2012.6233015

    work page arXiv 2012

  29. [32]

    Irith Pomeranz. 2024. Test Generation for Functionally Possible Subpaths. Trans. Comp. -Aided Des. Integ. Cir. Sys. 43, 12 (Dec. 2024), 4841 –4851. https://doi.org/10.1109/TCAD.2024.3396661

    arXiv 2024

  30. [33]

    Ecenur Ustun, Chenhui Deng, Debjit Pal, Zhijing Li, and Zhiru Zhang. 2020. Accurate operation delay prediction for FPGA HLS u sing graph neural networks. In Proceedings of the 39th International Conference on Computer-Aided Design (ICCAD '20). Association for Computing Machinery, New York, NY, USA, Article 87, 1–9. https://doi.org/10.1145/3400302.3415657

    work page arXiv 2020

  31. [34]

    Yanqing Zhang, Haoxing Ren, and Brucek Khailany. 2020. GRANNITE: graph neural network inference for transferable power estima tion. In Proceedings of the 57th ACM/EDAC/IEEE Design Automation Conference (DAC '20). IEEE Press, Article 60, 1 –6

    work page 2020

  32. [35]

    Le, James Laudon, Richard Ho, R oger Carpenter, and Jeff Dean

    Anna Goldie, Azalia Mirhoseini, Mustafa Yazgan, Joe Wenjie Jiang, Ebrahim Songhori, Shen Wang, Young -Joon Lee, Eric Johnson, Omkar Pathak, Azade Nova, Jiwoo Pak, Andy Tong, Kavya Srinivasa, William Hang, Emre Tuncer, Quoc V. Le, James Laudon, Richard Ho, R oger Carpenter, and Jeff Dean. 2024. Addendum: A graph placement methodology for fast chip design. ...

    work page doi:10.1038/s41586-024-08032-5 2024

  33. [36]

    Peng Cao, Yusen Qin, Guoqing He, Wenjie Ding, Xu Cheng, Zhanhua Zhang, and Yuyang Ye. 2025. An Optimization -Aware Pre-Routing Timing Prediction Framework Based on Multi-Modal Learning. IEEE Transactions on Computer -Aided Design of Integrated Circuits and S ystems. Article 1, 1 pages. https://doi.org/10.1109/TCAD.2025.3547806

    work page arXiv 2025

  34. [37]

    Bowen Wang, Guibao Shen, Dong Li, Jianye Hao, Wulong Liu, Yu Huang, Hongzhong Wu, Yibo Lin, Guangyong Chen, and Pheng Ann Heng. 2022. LHNN: lattice hypergraph neural network for VLSI congestion prediction. In Proceedings of the 59th ACM/IEEE Design Automat ion Conference (DAC '22). Association for Computing Machinery, New York, NY, USA, 1297 –1302. https:...

    arXiv 2022

  35. [38]

    Yuzhe Ma, Haoxing Ren, Brucek Khailany, Harbinder Sikka, Lijuan Luo, Karthikeyan Natarajan, and Bei Yu. 2019. High Performanc e Graph Convolutional Networks with Applications in Testability Analysis. In Proceedings of the 56th Annual Design Automation Conference 2019 (DAC '19). Association for Computing Machinery, New York, NY, USA, Article 18, 1 –6. http...

    work page arXiv 2019

  36. [39]

    Senling Wang, Shaoqi Wei, Hisashi Okamoto, Tatusya Nishikawa, Hiroshi Kai, Yoshinobu Higami, Hiroyuki Yotsuyanagi, Ruijun Ma, Tianmin g Ni, Hiroshi Takahashi, and Xiaoqing Wen. 2024. Test Point Selection for Multi-Cycle Logic BIST using Multivariate Temporal-Spatial GCNs. 2024 IEEE International Test Conference in Asia (ITC-Asia), 1–6. https://doi.org/10....

    work page arXiv 2024

  37. [40]

    Xavier Bresson and Thomas Laurent. 2017. Residual Gated Graph ConvNets. CoRR, arXiv:1711.07553. Retrieved from http://arxiv.org/abs/1711.07553

    Pith/arXiv arXiv 2017

  38. [41]

    Yunsheng Shi, Zhengjie Huang, Wenjin Wang, Hui Zhong, Shikun Feng, and Yu Sun. 2020. Masked Label Prediction: Unified Massage Passing Model 21 for Semi-Supervised Classification. CoRR, arXiv:2009.03509. Retrieved from https://arxiv.org/abs/2009.03509

    Pith/arXiv arXiv 2020

  39. [146]

    https://doi.org/10.1109/ATS.2006.261012

    work page arXiv 2006

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.