Pith. sign in

REVIEW 4 major objections 4 minor 48 references

DeepCell: Self-Supervised Multiview Fusion for Circuit Representation Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read DeepCell is the first representation-learning framework for post-mapping netlists; it fuses AIG and PM netlist views through masked circuit modeling to produce embeddings that cut ECO patch cost, gate count, runtime, and…

desk verdict DeepCell is a useful first PM netlist encoder with promising downstream results, but its central ablation conflates added Transformer capacity with multiview fusion, so the causality claim needs a capacity-matched control before it can be credited. read the letter →

arxiv 2502.06816 v2 pith:UATFX37Y submitted 2025-02-05 cs.LG cs.AI

Zhengyuan Shi , Chengyu Ma , Ziyang Zheng , Lingfeng Zhou , Hongyang Pan , Wentao Jiang , Fan Yang , Xiaoyan Yang
show 2 more authors
Zhufei Chu Qiang Xu
This is my paper · ORCID
classification cs.LGcs.AI
keywords circuitrepresentationlearningmultiviewfusionmaskmodelingpost-mappingnetlistengineeringchangeordertechnologymappingself-supervisedpretraininggraphneuralnetworks
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

DeepCell is a self-supervised framework that learns circuit representations by fusing two views of the same design: the pre-mapping And-Inverter Graph (AIG) and the post-mapping netlist of standard cells. The paper's central claim is that masking and reconstructing circuit cones across these two views, a task called Mask Circuit Modeling, yields unified embeddings that capture information no single view carries. The authors argue this is the first representation learner designed for post-mapping netlists, and they show it can be plugged into existing EDA tools. In functional engineering change order (ECO) tasks it lowers patch cost by 2.77%, gate count by 11.60%, and runtime by 25.63%; in technology mapping it cuts the area-delay product by about 15% to 16% on two technology libraries. If true, cross-stage EDA tasks would no longer need hand-built features for each design stage.

What carries the argument

Mask Circuit Modeling (MCM): a self-supervised pretraining task in which a k-hop cone of nodes in one circuit view is masked and the Transformer must reconstruct its functional embeddings using the unmasked tokens from both the AIG and the PM netlist. The masking preserves structural embeddings, so the model must recover function from structure plus the other view; the reconstruction loss is an L1 loss over the hidden functional embeddings. This mechanism is what forces the two encoders' outputs into a shared, information-rich space.

What would settle it

Train the full Transformer refinement on a PM netlist alone, with no AIG tokens, and compare prediction error against the reported multiview result; if the single-view Transformer matches the multiview gains, the fusion signal is not the cause.

Watch

Extended reading notes

Core claim

The paper's core discovery is that the technology-independent AIG view and the technology-dependent post-mapping view carry complementary information, and that a Transformer can be trained to bridge them with a masked-cone reconstruction objective. Starting from two GNN encoders pretrained to predict logic-1 probabilities under random simulation, DeepCell freezes one view's encoder and refines the other by masking a k-hop cone in one graph and reconstructing its functional embeddings using tokens from both graphs. The result is a single token sequence mixing AIG and PM netlist embeddings, refined by stacked Transformer blocks. The authors report that this multiview refinement improves representation quality for both views: prediction error drops by up to 28.03% for the AIG encoder and by roughly 9% for the PM encoder, and the learned embeddings transfer to ECO and technology mapping, where the model outperforms the baseline open-source EDA engine on cost, gate count, runtime, area, and delay.

Load-bearing premise

The load-bearing premise is that the reported gains come from fusing the two circuit views rather than from the extra Transformer refinement, because in the ablation the multiview condition adds both at once.

Editorial extensions

If this is right

  • DeepCell's embeddings can be dropped into existing ECO and technology-mapping flows as a plug-in, requiring only model inference time on top of the base EDA engine.
  • The multiview refinement improves both views: AIG encoders gain up to 28% lower prediction error, and PM encoders gain around 9%, suggesting the fusion works in both directions.
  • The model generalizes out-of-distribution to a technology library it was not pretrained on, keeping cell-type prediction accuracy above 90% after fine-tuning.
  • In technology mapping, better delay estimates from predicted loads also reduce area, because critical-path changes create room for area optimization in the mapper.

Reading between the lines

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

  • The reported ablations compare the full multiview model against a PM encoder without Transformer refinement, so the gains may partly reflect added model capacity; a cleaner test would refine the PM encoder with a Transformer using only PM tokens.
  • If the masked-cone objective generalizes, the same multiview fusion could be applied to other EDA stages, such as fusing RTL or Verilog views with netlists, or pre-layout with post-layout views.
  • The approach suggests a general recipe: pretrain per-view GNN encoders with simulation-based supervision, then use a masked-reconstruction Transformer to align views, which could transfer to other graph-structured engineering artifacts beyond circuits.
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.

Referee Report

4 major / 4 minor

Summary. DeepCell proposes a self-supervised multiview representation learning framework for circuit netlists, fusing And-Inverter Graph (AIG) and post-mapping (PM) netlist views. The method trains GNN encoders on each view, concatenates their node embeddings, masks circuit cones, and uses a Transformer to reconstruct masked functional embeddings (Mask Circuit Modeling, MCM). The authors evaluate the learned representations on logic-1 probability prediction (PE) and reconstruction error (RE), and demonstrate two downstream applications: functional ECO, where DeepCell prunes candidate signals in ABC's 'runeco' flow, and technology mapping, where a finetuned AIG encoder predicts cell types to improve delay estimation in ABC's '&nf' command. The paper claims state-of-the-art accuracy and reconstruction quality for PM netlist representation learning, reporting reductions in ECO patch cost and technology-mapping area-delay product.

Significance. If the central claim is established, DeepCell would be a worthwhile contribution: it is among the first attempts to learn general-purpose representations of post-mapping netlists, it naturally combines technology-independent and technology-dependent views, and the plug-in integration into ABC gives it concrete practical value for cross-stage EDA tasks. The paper also contains useful experimental assets: pretraining over multiple open-source technology libraries, an out-of-distribution test set, and two realistic downstream benchmarks. However, the evidence for the core novelty -- that multiview fusion itself improves representation learning -- is currently weakened by an ablation that adds model capacity together with the multiview signal, and by an evaluation metric (RE) that measures reconstruction of the model's own learned embeddings rather than any external ground truth. The downstream results are promising but are reported without variance and with some inconsistencies in timeout handling.

major comments (4)
  1. [Section IV-B2/IV-B3, Tables III and IV] The w/ vs w/o multiview comparison is confounded by model capacity. In the w/o multiview condition, the model consists only of the GNN encoder, while the w/ multiview condition adds both the Transformer refinement (Eq. (7), Fig. 1) and the second view. The reported PE reductions (e.g., 28.03% for DeepGate2 in Table IV, 9.41% for Conv. Sum in Table III) therefore conflate the effect of the 4-layer, 8-head Transformer with the effect of multiview fusion. To isolate the multiview signal, the authors should add a control that uses the same Transformer refinement under a single-view MCM objective (masking and reconstructing from only the PM view or only the AIG view), while keeping encoder architecture fixed. In addition, none of the PE numbers in Tables III and IV are accompanied by variance across random seeds; given the modest size of most gains (2.7-9.9%), this is necessary to determine whether the differences are meaningful.
  2. [Section III-D, Eqs. (10) and (14); Abstract] The 'reconstruction quality' claim is self-referential. The MCM loss in Eq. (10) regresses the functional embeddings h_f^P that are themselves produced by the same GNN encoder, and the RE metric in Eq. (14) measures the distance to those learned embeddings rather than to any external ground truth. Consequently, the abstract's assertion that DeepCell sets 'new benchmarks in both predictive accuracy and reconstruction quality' is not supported by an independent reconstruction benchmark. The PE metric (Eq. (13)) is supervised by random-simulation logic-1 probabilities and is therefore more meaningful, but the paper should either rephrase the reconstruction-quality claim or validate it with a downstream task that does not use the model's own embeddings as targets.
  3. [Section V-D, Tables V and VI] The ECO results contain inconsistencies and unclear aggregation. In Table VI, units 6, 10, 19 are shown as timeouts ('-') even in the w/ DeepCell condition, yet Table V reports that unit 6 is solved by w/ DeepCell in 538.06 s and unit 11 in 2379.78 s. It is not explained whether Table VI uses the same time limit as Table V or what exactly the 'time(s)' and 'overall(s)' columns represent for the w/ condition. The footnote excludes 'easy cases' and 'hard cases' from the geomean and reduction calculations, so the reported reductions of 2.77%, 11.60%, and 25.63% are not straightforward averages over the full benchmark. The authors should give a precise protocol for handling timeouts, report the number of solved units in each condition, and clarify the relationship between Tables V and VI.
  4. [Section VI-D, Table VII] The narrative in Section VI-D overstates the delay results. The text says 'DeepCell improves delay performance compared to the baseline engine,' but several rows of Table VII show delay increases in the w/ DeepCell column (e.g., adder: 7,314.47 to 7,362.27 ps; arbiter: 1,641.73 to 1,930.43 ps; div: 211,175.92 to 242,915.17 ps; sqrt: 406,476.56 to 420,115.69 ps under Skywater130nm). The valid and still useful claim is the aggregate area-delay product (ADP) reduction of 15.02% and 16.00%, which appears to be driven by area savings rather than by delay improvements. The wording should be revised to reflect the actual trade-off, and the authors should report the distribution of per-circuit changes rather than only geomeans.
minor comments (4)
  1. [Eq. (9)] The index in Eq. (9) is written as i in V_P, but this equation defines the AIG encoder loss and should refer to nodes in V_A.
  2. [Section III-D] The sentence 'All the selected masked nodes p is in set S' has a subject-verb agreement error and should read 'All selected masked nodes p are in set S.'
  3. [Section IV-B1, Table II] The claim that 'masking a smaller number of nodes consistently reduces both the RE and PE' is not strictly supported for RE: for k=4, the RE at θ=0.05 (0.0097) is lower than at θ=0.01 (0.0099), so the relationship is not monotonic over the tested range.
  4. [Section VI-B, reference [40]] The text states that the Transformer Blocks are implemented using 'Linear Transformer' and cites reference [40], which is the Reformer paper. These are different efficient-attention mechanisms; the authors should cite the correct linear-transformer work or change the description.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-referential reconstruction metric; downstream benchmarks keep the central claim independent.

  1. self definitional [Section III-E, Eq. (10); Section IV-A2, Eq. (14)]
    "we define the MCM training loss in Eq. (10), where the model is trained to recover the functional embeddings of the masked cells. Lmcm = L1Loss(H P i , HP ′ i ) (10) ... the average reconstruction error (RE) is defined as Eq. (14) ... RE = 1/|M(p)| Σ i∈M(p) |hf P i − hf P ′ i |, p∈ S (14)"

    The reconstruction-quality metric RE is the same L1 error as the MCM pretraining loss Lmcm, and both compare the model's own functional embeddings hf^P_i (produced by the unmasked PM encoder) with the Transformer's reconstruction hf^{P'}_i. Thus the reported "reconstruction quality" is the training objective itself by construction; a low RE shows that the model can reproduce its own internal embeddings, not that it has matched an external ground-truth circuit property. This makes the abstract's "reconstruction quality" benchmark self-referential. However, the downstream ECO and technology-mapping evaluations use external objectives (patch cost, gate count, runtime, area, delay), so the central multiview representation claim does not reduce to this metric.

full rationale

The paper's derivation chain is largely self-contained and anchored to external benchmarks. Stage-1 pretraining (Eqs. 8-9) supervises encoders against logic-1 probabilities from random simulation, an external circuit-level label; Stage-2 MCM is a standard self-supervised pretext task. The only concrete reduction I can exhibit is that the RE metric (Eq. 14) is identical to the MCM loss (Eq. 10), so the paper's "reconstruction quality" claim is a self-consistency measure rather than independent evidence. This is a minor circular element because the central novelty claim is not supported by RE alone; it is supported by the ECO comparison against the ICCAD'17 champion flow (Table VI) and technology-mapping comparison against ABC's &nf (Table VII), both of which use external cost metrics. The w/ vs w/o multiview ablation (Tables III-IV) conflates adding the Transformer refinement with adding cross-view information, but that is an experimental confound rather than an equation-level circularity and is not counted here. Self-citations, such as using DeepGate2 as the AIG encoder or citing ForgeEDA/DeepCircuitX datasets, are not load-bearing: the paper evaluates alternative AIG encoders (PolarGate, HOGA, DeepGate3, GCN) and alternative PM aggregators, and the cited prior works are independently published. No uniqueness theorem is imported from the authors' prior work, and no known result is merely renamed. Overall, the central derivation is independent of its inputs, with one self-referential auxiliary metric.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

DeepCell introduces no new physical entities or mediators. Its dependency chain is methodological: logic-1 probability supervision from DeepGate, the DeepGate2 AIG encoder, an unnamed commercial synthesis tool, and the assumption that self-reconstruction of learned embeddings is a useful pretraining signal. The most fragile premise is the ablation design, which does not separate multiview fusion from added Transformer capacity.

free parameters (5)
  • mask ratio theta = 0.05
    Selected via hyperparameter search on validation PE/RE in Table II; tuning on the same data as the reported results makes this a fitted value.
  • masked hop size k = 4
    Chosen from {4, 6} based on lower reconstruction error in Table II.
  • loss weights wprob, wmcm = 1 each
    Set by default without reported ablation (Section III-E).
  • embedding dimension = 128
    Structural and functional embeddings fixed at 128 dimensions without sensitivity analysis.
  • node feature dimension D = 64
    Chosen to fit truth tables of cells with up to 6 inputs (Eq. 1); fixed without ablation.
assumptions (5)
  • domain assumption Logic-1 probability under random simulation is a useful self-supervisory signal for circuit encoders.
    Adopted from DeepGate [3]; used as the Stage 1 pretraining target (Section III-E, Eqs. 8-9). The paper does not question this inherited choice.
  • domain assumption DeepGate2 is an adequate frozen AIG encoder for the fusion framework.
    DeepGate2 [5] is used as the default AIG backbone (Section III-C); it is a self-cited model from the same research group. Its quality directly affects the PM encoder gains.
  • ad hoc to paper Masked autoencoding of the model's own functional embeddings transfers to downstream EDA tasks.
    The MCM objective (Eq. 10) reconstructs learned embeddings, not ground-truth circuit properties. The paper's only evidence that this transfer works is the downstream finetuning results.
  • domain assumption PM netlists generated by the commercial synthesis tool are representative of real designs.
    Section IV-A1 states a commercial tool generates post-mapping netlists, but the tool is unnamed. The ID/OOD split across libraries is the only check on this assumption.
  • ad hoc to paper The w/o multiview baseline isolates the effect of multiview fusion.
    Tables III-IV compare w/ multiview (GNN plus Transformer) vs w/o multiview (GNN only). This assumes the Transformer does not itself contribute the reported gains, which is not controlled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepCell: Self-Supervised Multiview Fusion for Circuit Representation Learning." pith.science (2026). https://pith.science/paper/UATFX37Y

@misc{pith2026250206816,
  author       = {Pith},
  title        = {Pith review of: DeepCell: Self-Supervised Multiview Fusion for Circuit Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UATFX37Y}},
  note         = {Machine review of arXiv:2502.06816}
}
read the original abstract

We introduce DeepCell, a novel circuit representation learning framework that effectively integrates multiview information from both And-Inverter Graphs (AIGs) and Post-Mapping (PM) netlists. At its core, DeepCell employs a self-supervised Mask Circuit Modeling (MCM) strategy, inspired by masked language modeling, to fuse complementary circuit representations from different design stages into unified and rich embeddings. To our knowledge, DeepCell is the first framework explicitly designed for PM netlist representation learning, setting new benchmarks in both predictive accuracy and reconstruction quality. We demonstrate the practical efficacy of DeepCell by applying it to critical EDA tasks such as functional Engineering Change Orders (ECO) and technology mapping. Extensive experimental results show that DeepCell significantly surpasses state-of-the-art open-source EDA tools in efficiency and performance.

Figures

Figures reproduced from arXiv: 2502.06816 by the authors.

Figure 1
Figure 1. Overview of DeepCell illustrating the multiview fusion of AIG and PM netlist embeddings [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. To be specific, given a PM netlist G P and the corresponding AIG G A, we randomly select node p ∈ GP and extract a k-hop predecessors around node p, denoted as the Masked Hop M(p). All nodes i ∈ M(p) are considered as masked, and their functional embeddings hfi are replaced by the learnable masked token hm, while their structural embeddings hsi are preserving. As a result, we have M + N tokens (see Eq. (6)) in total… view at source ↗
Figure 2
Figure 2. Mask circuit modeling of PM netlists H P i = {hsP i , hf P i }, i ∈ VP , i /∈ S H P i = {hsP i , hm}, i ∈ VP , i ∈ S H A j = {hsA j , hf A j }, j ∈ VA (6) Then, we process these M +N tokens using a Transformer model T . Formally, we define the input and output of the Transformer model as below. {H A′ , H P ′ } = T ({H A , H P }) (7) Similarly, when the DeepCell framework is applied to refine the AIG embeddings, the … view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: DeepCell finetuning for functional ECO each representing the probability that a given cell should be included in the candidate signal set. C. Experiment Settings Our model is equipped into ABC [17] as a plug-in and integrated into the command ‘runeco’, which is an effi…
Figure 4
Figure 4. Figure 4: DeepCell finetuning for technology mapping [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 38 canonical work pages

  1. [1]

    Large circuit models: opportunities and challenges,

    L. Chen, Y . Chen, Z. Chu, W. Fang, T.-Y . Ho, R. Huang, Y . Huang, S. Khan, M. Li, X. Li et al. , “Large circuit models: opportunities and challenges,” Science China Information Sciences , vol. 67, no. 10, pp. 1–42, 2024

    work page 2024

  2. [2]

    Deepcircuitx: A comprehensive repository-level dataset for rtl code understanding, generation, and ppa analysis,

    Z. Li, C. Xu, Z. Shi, Z. Peng, Y . Liu, Y . Zhou et al. , “Deepcircuitx: A comprehensive repository-level dataset for rtl code understanding, generation, and ppa analysis,” arXiv preprint arXiv:2502.18297 , 2025

    arXiv 2025

  3. [3]

    Deepgate: Learning neural representations of logic gates,

    M. Li, S. Khan, Z. Shi, N. Wang, H. Yu, and Q. Xu, “Deepgate: Learning neural representations of logic gates,” in Proceedings of the 59th ACM/IEEE Design Automation Conference , 2022, pp. 667–672

    2022

  4. [4]

    Functionality matters in netlist representation learning,

    Z. Wang, C. Bai, Z. He, G. Zhang, Q. Xu, T.-Y . Ho, B. Yu, and Y . Huang, “Functionality matters in netlist representation learning,” in Proceedings of the 59th ACM/IEEE Design Automation Conference, 2022, pp. 61–66

    work page 2022

  5. [5]

    Deepgate2: Functionality-aware circuit representation learning,

    Z. Shi, H. Pan, S. Khan, M. Li, Y . Liu, J. Huang, H.-L. Zhen, M. Yuan, Z. Chu, and Q. Xu, “Deepgate2: Functionality-aware circuit representation learning,” in 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD) . IEEE, 2023, pp. 1–9

    work page 2023

  6. [6]

    Gamora: Graph learning based symbolic reasoning for large-scale boolean networks,

    N. Wu, Y . Li, C. Hao, S. Dai, C. Yu, and Y . Xie, “Gamora: Graph learning based symbolic reasoning for large-scale boolean networks,” in 2023 60th ACM/IEEE Design Automation Conference (DAC) . IEEE, 2023, pp. 1–6

    work page 2023

  7. [7]

    Less is More: Hop-Wise Graph Attention for Scalable and Generalizable Learning on Circuits

    C. Deng, Z. Yue, C. Yu, G. Sarar, R. Carey, R. Jain, and Z. Zhang, “Less is more: Hop-wise graph attention for scalable and generalizable learning on circuits,” arXiv preprint arXiv:2403.01317 , 2024

    work page Pith review arXiv 2024

  8. [8]

    Deepgate3: Towards scalable circuit representation learning,

    Z. Shi, Z. Zheng, S. Khan, J. Zhong, M. Li, and Q. Xu, “Deepgate3: Towards scalable circuit representation learning,” in Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design , 2024, pp. 1–9

    work page 2024

Show all 48 references
  1. [9]

    Deepseq: Deep sequential circuit learning,

    S. Khan, Z. Shi, M. Li, and Q. Xu, “Deepseq: Deep sequential circuit learning,” in 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 2024, pp. 1–2

    2024

  2. [10]

    Polargate: Breaking the functionality representation bottleneck of and-inverter graph neural network,

    J. Liu, J. Zhai, M. Zhao, Z. Lin, B. Yu, and C. Shi, “Polargate: Breaking the functionality representation bottleneck of and-inverter graph neural network,” in 2024 IEEE/ACM International Conference on Computer Aided Design (ICCAD) . IEEE, 2024, pp. 1–9

    2024

  3. [11]

    Deeptpi: Test point insertion with deep reinforcement learning,

    Z. Shi, M. Li, S. Khan, L. Wang, N. Wang, Y . Huang, and Q. Xu, “Deeptpi: Test point insertion with deep reinforcement learning,” in2022 IEEE International Test Conference (ITC) . IEEE, 2022, pp. 194–203

    2022

  4. [12]

    Lsoracle: A logic synthesis framework driven by arti- ficial intelligence,

    W. L. Neto, M. Austin, S. Temple, L. Amaru, X. Tang, and P.-E. Gaillardon, “Lsoracle: A logic synthesis framework driven by arti- ficial intelligence,” in 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD) . IEEE, 2019, pp. 1–6

    2019

  5. [13]

    Rtlrewriter: Methodologies for large models aided rtl code optimization,

    X. Yao, Y . Wang, X. Li, Y . Lian, R. Chen, L. Chen, M. Yuan, H. Xu, and B. Yu, “Rtlrewriter: Methodologies for large models aided rtl code optimization,” in Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design , 2024, pp. 1–7

    2024

  6. [14]

    Circuitfusion: multimodal circuit representation learning for agile chip design,

    W. Fang, S. Liu, J. Wang, and Z. Xie, “Circuitfusion: multimodal circuit representation learning for agile chip design,” in The Thirteenth International Conference on Learning Representations , 2025

    2025

  7. [15]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    2018 arXiv

  8. [16]

    Efficient computation of eco patch functions,

    A. Q. Dao, N.-Z. Lee, L.-C. Chen, M. P.-H. Lin, J.-H. R. Jiang, A. Mishchenko, and R. Brayton, “Efficient computation of eco patch functions,” in Proceedings of the 55th Annual Design Automation Conference, 2018, pp. 1–6

    2018

  9. [17]

    Abc: A system for sequential synthesis and verification,

    A. Mishchenko et al. , “Abc: A system for sequential synthesis and verification,” URL http://www. eecs. berkeley. edu/alanmi/abc , vol. 17, 2007

    2007

  10. [18]

    Logic optimization meets sat: A novel framework for circuit-sat solving,

    Z. Shi, T. Tang, J. Zhu, S. Khan, H.-L. Zhen, M. Yuan, Z. Chu, and Q. Xu, “Logic optimization meets sat: A novel framework for circuit-sat solving,” 2025. [Online]. Available: https://arxiv.org/abs/2403.19446

    2025 arXiv

  11. [19]

    Deepgate4: Efficient and effective representation learning for circuit design at scale,

    Z. Zheng, S. Huang, J. Zhong, Z. Shi, G. Dai, N. Xu, and Q. Xu, “Deepgate4: Efficient and effective representation learning for circuit design at scale,” arXiv preprint arXiv:2502.01681 , 2025

    2025 arXiv

  12. [20]

    Virtually every asic ends up an fpga,

    J. Jaeger, “Virtually every asic ends up an fpga,” EE Times, 2007

    2007

  13. [21]

    Match and replace: A functional eco engine for multierror circuit rectification,

    S.-L. Huang, W.-H. Lin, P.-K. Huang, and C.-Y . Huang, “Match and replace: A functional eco engine for multierror circuit rectification,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 32, no. 3, pp. 467–478, 2013

    2013

  14. [22]

    Engineering change order for combinational and sequential design rectification,

    J.-H. R. Jiang, V . N. Kravets, and N.-Z. Lee, “Engineering change order for combinational and sequential design rectification,” in 2020 Design, Automation & Test in Europe Conference & Exhibition (DATE) . IEEE, 2020, pp. 726–731

    2020

  15. [23]

    Learning to automate the design updates from observed engineering changes in the chip development cycle,

    V . N. Kravets, J.-H. R. Jiang, and H. Riener, “Learning to automate the design updates from observed engineering changes in the chip development cycle,” in 2020 Design, Automation & Test in Europe Conference & Exhibition (DATE) . IEEE, 2020, pp. 738–743

    2020

  16. [24]

    Feep: Functional eco synthesis with efficient patch minimization,

    Y . Liu, Y . Zhang, Q. Zhang, R. Chen, and Y . Li, “Feep: Functional eco synthesis with efficient patch minimization,” in 2023 IEEE 5th International Conference on Artificial Intelligence Circuits and Systems (AICAS). IEEE, 2023, pp. 1–5

    2023

  17. [25]

    Reducing structural bias in technology mapping,

    S. Chatterjee, A. Mishchenko, R. K. Brayton, X. Wang, and T. Kam, “Reducing structural bias in technology mapping,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , vol. 25, no. 12, pp. 2894–2903, 2006

    2006

  18. [26]

    An enhanced resubstitution algorithm for area-oriented logic optimization,

    A. Costamagna, A. Mishchenko, S. Chatterjee, and G. De Micheli, “An enhanced resubstitution algorithm for area-oriented logic optimization,” in 2024 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2024, pp. 1–5

    2024

  19. [27]

    Logical effort: Designing for speed on the back of an envelope,

    I. E. Sutherland and R. F. Sproull, “Logical effort: Designing for speed on the back of an envelope,” IEEE Advanced Research in VLSI , vol. 116, p. 270, 1991

    1991

  20. [28]

    Technology mapping with boolean matching, supergates and choices,

    A. Mishchenko, S. Chatterjee, R. Brayton, X. Wang, and T. Kam, “Technology mapping with boolean matching, supergates and choices,” Berkeley Logic Synthesis and Verification Group , 2005

    2005

  21. [29]

    Maptune: Advancing asic technology mapping via reinforcement learning guided library tuning,

    M. Liu, D. Robinson, Y . Li, and C. Yu, “Maptune: Advancing asic technology mapping via reinforcement learning guided library tuning,” in Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design, 2024, pp. 1–10

    2024

  22. [30]

    Timing-driven technology mapping approximation based on reinforcement learning,

    Y . Ye, T. Chen, Y . Gao, H. Yan, B. Yu, and L. Shi, “Timing-driven technology mapping approximation based on reinforcement learning,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024

    2024

  23. [31]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

    2016 arXiv

  24. [32]

    Forgeeda: A comprehensive multimodal dataset for advancing eda,

    Z. Shi, Z. Li, C. Ma, Y . Zhou, Z. Zheng, J. Liu, H. Pan, L. Zhou, K. Li, J. Zhu et al., “Forgeeda: A comprehensive multimodal dataset for advancing eda,” arXiv preprint arXiv:2505.02016 , 2025

    2025 arXiv

  25. [33]

    Skywater open source pdk

    Google. Skywater open source pdk. 2020. [Online]. Available: https://github.com/google/skywater-pdk.git

    2020

  26. [34]

    Free45 pdk. 2023. [Online]. Available: https: //github.com/The-OpenROAD-Project-Attic/OpenROAD-Utilities/ tree/master/TimerCalibration/Free45PDK

    2023

  27. [35]

    Globalfoundries gf180mcu open source pdk

    Google. Globalfoundries gf180mcu open source pdk. 2022. [Online]. Available: https://github.com/google/gf180mcu-pdk.git

    2022

  28. [36]

    Nangate-design optimization company. 2009. [Online]. Available: http://www.nangate.com/index.php

    2009

  29. [37]

    Learning a sat solver from single-bit supervision,

    D. Selsam, M. Lamm, B. B ¨unz, P. Liang, L. de Moura, and D. L. Dill, “Learning a sat solver from single-bit supervision,” arXiv preprint arXiv:1802.03685, 2018

    2018 arXiv

  30. [38]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903, 2017

    2017 arXiv

  31. [39]

    A robust functional eco engine by sat proof minimization and interpolation techniques,

    B.-H. Wu, C.-J. Yang, C.-Y . Huang, and J.-H. R. Jiang, “A robust functional eco engine by sat proof minimization and interpolation techniques,” in 2010 IEEE/ACM International Conference on Computer- Aided Design (ICCAD) . IEEE, 2010, pp. 729–734

    2010

  32. [40]

    Reformer: The efficient trans- former,

    N. Kitaev, L. Kaiser, and A. Levskaya, “Reformer: The efficient trans- former,” in International Conference on Learning Representations, 2020

    2020

  33. [41]

    2017 iccad cad contest problem a: Resource-aware patch generation,

    C. Huang, C. Hsu, and C. Wu, “2017 iccad cad contest problem a: Resource-aware patch generation,” in Proc. ICCAD, 2017

    2017

  34. [42]

    Combinational profiles of sequential benchmark circuits,

    F. Brglez, D. Bryan, and K. Kozminski, “Combinational profiles of sequential benchmark circuits,” in 1989 IEEE International Symposium on Circuits and Systems (ISCAS) . IEEE, 1989, pp. 1929–1934

    1989

  35. [43]

    Characteristics of the itc’99 benchmark circuits,

    S. Davidson, “Characteristics of the itc’99 benchmark circuits,” in IEEE International Test Synthesis Workshop (ITSW) , 1999, p. 87

    1999

  36. [44]

    Iwls 2005 benchmarks,

    C. Albrecht, “Iwls 2005 benchmarks,” in International Workshop for Logic Synthesis (IWLS) , vol. 9, 2005

    2005

  37. [45]

    Boolean matching for large libraries,

    U. Hinsberger and R. Kolla, “Boolean matching for large libraries,” in Proceedings of the 35th annual Design Automation Conference , 1998, pp. 206–211

    1998

  38. [46]

    Delay-optimal technology mapping by dag covering,

    Y . Kukimoto, R. K. Brayton, and P. Sawkar, “Delay-optimal technology mapping by dag covering,” in Proceedings Of The 35th Annual Design Automation Conference, 1998, pp. 348–351

    1998

  39. [47]

    The epfl combinational benchmark suite,

    L. Amar ´u, P.-E. Gaillardon, and G. De Micheli, “The epfl combinational benchmark suite,” in Proceedings of the 24th International Workshop on Logic & Synthesis (IWLS) , 2015

    2015

  40. [48]

    Practical boolean decomposition for delay-driven lut mapping,

    A. T. Calvino, A. Mishchenko, G. De Micheli, and R. Brayton, “Practical boolean decomposition for delay-driven lut mapping,” arXiv preprint arXiv:2406.06241, 2024

    2024 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.