acceptodds
Under review as a conference paper at ICLR 2027

Can Ground-Capacitance Predictors Generalize Across Circuit Domains?

Abstract

Producing circuit layouts requires substantial engineering effort, and failed post- layout verification can trigger further layout, extraction, and simulation iterations. These costs, together with restricted access to circuit intellectual property, make labeled parasitic data scarce when entering a new circuit regime. Learning-based prediction could provide earlier electrical feedback, but its usefulness in this cold-start setting depends on reusing data from other circuit domains. We study target-label-free, node-level ground-capacitance prediction across Digital, Analog, and SRAM circuits, where similar graph structures can carry different physical contexts and loading behavior. We propose XRC-Former, a graph Transformer that preserves circuit context through domain-sensitive encoding, reuses structural prototypes through shared memory, and aligns only a designated invariant branch while retaining a non-aligned residual for prediction. Under leave-one-domain- out evaluation, using source labels for training and source validation for model selection, XRC-Former achieves the best target-domain R2 among 14 task-adapted baselines. It remains positive in the difficult Analog-target setting, where most baselines obtain negative R2. Ablations, embedding diagnostics, and supervised within-domain references characterize the roles and scope of the proposed transfer mechanism. The results support cross-domain GC prediction as a step toward reusable parasitic screening before the final layout stage, rather than a replacement for signoff verification.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
domain generalization, cross-domain transfer learning, graph representation learning, graph transformers, circuit graph learning, electronic design automation, parasitic capacitance prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2REbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies zero-shot ground-capacitance prediction across three circuit domains: Digital, Analog, and SRAM. The authors use a leave-one-domain-out (LODO) setting, training on two domains and evaluating on the third without target labels.

XRC-Former combines three components: (1) domain-sensitive positional encoding (DSPE) using topology, geometry, and physics information, (2) a shared prototype memory bank for retrieving structural patterns, and (3) invariant and domain-conditioned representations, with MMD alignment applied to the invariant branch.

Experiments against 14 baselines show the best target R² across all three domains. The Analog results are particularly notable, as most competing methods obtain negative R². The paper also includes ablations, representation analysis, and few-shot adaptation experiments.

Critical strengths

(1) Interesting problem and evaluation setting. Zero-shot prediction across circuit families is a practically relevant problem. The LODO setup is appropriate, and the Analog fold is particularly challenging, with most baselines performing worse than predicting the target-domain mean. The results highlight a genuine difficulty in transferring circuit representations across domains.

(2) Careful documentation of the evaluation protocol. Appendix B clearly separates LODO evaluation from few-shot adaptation and specifies source-only checkpoint selection. I also appreciate that Appendix D.1 acknowledges the technology confound, while Appendix E.1 avoids treating the within-domain reference as a strict upper bound. These clarifications help establish the intended scope of the results.

Critical weaknesses

(1) The label normalization protocol is unclear, and this directly affects the zero-shot claim. Appendix D.2 states that label_norm_mode = legacy is used and metrics are computed in normalized space without inverse transformation. However, I could not find whether normalization statistics are computed separately for each domain or from source data alone.

This is important because the raw capacitance distributions differ substantially across domains (Tables D.1–D.3). If target-domain label statistics are used, the evaluation is no longer entirely target-label-free. If normalization is source-derived, the large negative R² values of several baselines may partly reflect differences in target scale rather than an inability to learn transferable structure. Either way, this needs clarification before interpreting Table 1.

The statistical reporting also needs attention. Appendix D.4 mentions multiple seeds but does not specify how many were used for Table 1. The Analog result of 0.4245 ± 0.1991 has considerable variability, and the Analog corpus contains some very small cases (Table D.2). Without the number of runs or per-run results, it is difficult to judge how reliable the improvement is.

(2) The main architectural contributions are not sufficiently justified under the LODO setting. In Eq. (3), the geometry and physics encoders are domain-conditioned. But the target domain is unseen during training, so which parameters are used for its encoding? Table F1 mentions zero-fallback for Analog and SRAM in GeoPE, but the corresponding treatment of PhysPE is unclear.

This is especially important for Analog, where removing DSPE decreases R² from 0.4245 to −0.7942 (Figure 3). If its geometry branch is zeroed for the unseen domain, what exactly accounts for this large contribution?

I also have concerns about the invariant/domain-conditioned decomposition in Eq. (8). Since the two projected representations are simply added before prediction, the forward computation is equivalent to applying a single linear projection, (Winv+Wdom)h(L)(W_{\mathrm{inv}}+W_{\mathrm{dom}})h^{(L)}. The distinction is introduced through MMD regularization, rather than a separate architectural mechanism for preserving domain-specific information.

Finally, the ablation results on Analog are unusually sensitive. Removing any one of DSPE, memory, or alignment pushes R² below zero. In particular, removing alignment alone costs approximately 0.67 R², despite alignment being applied to just one source-domain pair with λalign=0.1\lambda_{\mathrm{align}}=0.1. The components may indeed be strongly interdependent, but the paper needs to explain these large, non-additive effects.

Questions for authors

(1) How exactly are the ground-capacitance labels normalized? Are the statistics computed separately for each domain or using source domains only? If target statistics are used, how is this compatible with the zero-target-label setting? I would like to see Table 1 evaluated with source-derived normalization.

(2) Under LODO, how are ϕDgeo\phi_D^{\mathrm{geo}} and ϕDphys\phi_D^{\mathrm{phys}} handled for the unseen target domain? In particular, if Analog uses a zero-fallback GeoPE, why does removing DSPE cause such a large drop in Figure 3?

(3) How many random seeds were used for Table 1? Could the authors provide the per-run results or paired significance tests, especially for Analog, where the reported variance is relatively large?

(4) Could the authors include an ablation using shared geometry and physics encoders across domains, rather than removing DSPE entirely? This would help determine whether the improvement comes from circuit-aware positional encoding or the proposed domain conditioning.

(5) Appendix D.1 notes that the three circuit domains also use different process technologies (45 nm, 180 nm BCD, and 28 nm). Is there a way to separate the effect of circuit domain from process technology, perhaps through within-technology transfer or cross-design Digital-to-Digital evaluation?

(6) Could the authors report the memory-bank diagnostics, particularly active prototype counts and assignment entropy? Since the memory loss encourages node states to approach their retrieved prototypes, while those prototypes are updated from the same states, I am concerned about possible prototype collapse.

Minor Comments

(A) CirGPS and CircuitGPS are used inconsistently in Section 2.1, Table 1, and Table E3. (B) Typo: "propsoed" in Section 2.1. (C)Table D.2 lists identical statistics for case11 and case17. Please check whether this is a duplication. (D) Section 1 suggests the method extends to coupling capacitance and resistance, although Section 2.1 acknowledges that these require different output structures. I would avoid making this claim without further evidence.

The claim of consistently superior performance should be qualified by the metric: Table E2 reports better Digital MAE for CircuitGPS, although XRC-Former leads in R².

2 comments

Thank you for the detailed review. We address the implementation, ablations, and scope below.

1. Fixed label normalization and reproducibility

The legacy function uses fixed transforms without fitting target-domain means, SDs, minima, or maxima:

TA(C)=clip⁡(log⁡10(max⁡(C,10−30)1020)6,0,1),T_A(C)=\operatorname{clip}\left(\frac{\log_{10}(\max(C,10^{-30})10^{20})}{6},0,1\right), TD(C)=clip⁡(log⁡10(max⁡(C,10−30)106)6,0,1).T_D(C)=\operatorname{clip}\left(\frac{\log_{10}(\max(C,10^{-30})10^6)}{6},0,1\right).

Analog/Digital inputs use F/pF, respectively. These are domain-dependent label transforms, with a 1/31/3 offset for equal physical capacitance before clipping. We will explicitly distinguish unit conversion from label transformation and specify the clipping and metric space.

All experiments in our updated results use five random seeds. We will document the final loader/version, units, initialization, split seed, and checkpoint-selection rules, with each mean and SD computed from the same five runs.

2. Shared projections and held-out mappings

WinvW_{\mathrm{inv}} and WdomW_{\mathrm{dom}} are globally shared projections. They are separate from the domain-indexed GeoPE/PhysPE MLPs. Under fresh source-only training, an enabled held-out PhysPE mapping retains its initialization; it is not a source-trained domain-specific encoder. We will distinguish these parameter groups and explicitly state the held-out mapping's initialization and inference behavior.

3. Eq. (8) and asymmetric alignment

We agree that:

hinv+hdom=(Winv+Wdom)h+binv+bdom.h_{\mathrm{inv}}+h_{\mathrm{dom}}=(W_{\mathrm{inv}}+W_{\mathrm{dom}})h+b_{\mathrm{inv}}+b_{\mathrm{dom}}.

The distinction lies in the training objective: MMD regularizes hinvh_{\mathrm{inv}}, while prediction uses both branches. This imposes different regularization on the two paths. We will describe them as aligned/non-aligned branches and avoid implying a formal identifiability guarantee.

4. Positional-encoding ablations and memory usage

The Analog-target diagnostic uses five seeds (0, 1, 2, 42, 123), split_seed=42, 80 epochs, patience 15, and source-validation checkpointing. Values are mean ± population SD:

VariantMAERMSER2R^2
Full.1989±.0005.2212±.0062-.1440±.0646
TopoPE-only.2021±.0011.2434±.0043-.3844±.0491
w/o DSPE.2090±.0052.2389±.0065-.3348±.0741
w/o memory.1984±.0005.2193±.0052-.1247±.0534
w/o alignment.1978±.0025.2194±.0043-.1251±.0439

TopoPE-only retains shared Laplacian/random-walk encodings and disables GeoPE/PhysPE. Within this diagnostic, Full reduces MAE by 1.6% and RMSE by 9.1% versus TopoPE-only. The paired R2R^2 gain is 0.2405±0.02190.2405\pm0.0219 (population SD), with higher R2R^2 in all five seeds. This comparison concerns the GeoPE/PhysPE paths jointly.

Separately, five-seed real-three-domain raw-data runs with K=64K=64 provide source-training hard-assignment statistics at source-validation-selected epochs. For the Analog-target fold, occupied slots are 54.8±1.154.8\pm1.1, normalized entropy H/ln⁡64=.864±.017H/\ln64=.864\pm.017, and top-slot share (8.01±1.95)%(8.01\pm1.95)\% (sample SD). These training-side statistics support distributed slot usage in those records; they are separate from Table 1 and are not frozen-target diagnostics.

5. PDK, data reporting, and evaluation scope

Parasitic labels are costly to obtain and fragmented across circuit families and PDKs. This motivates our application: reusing available circuit data to support transfer when target labels are scarce. The evaluated setting combines family and process shifts; we will make that scope explicit, with common-PDK controls as future work.

Thank you for identifying the case11/case17 duplication. It is a typographical error in the dataset table: the entries refer to one case, which should appear only once. We will correct the duplicated entry and corresponding description.

CircuitGPS/CircuitGCL achieve lower Digital MAE in Table E2. MAE, RMSE, and R2R^2 provide complementary views, and overall transfer should be assessed across target domains. We will frame our contribution around reusing scarce circuit labels for transferable prediction and qualify comparisons by metric and protocol. Formal evaluation focuses on ground capacitance; coupling-capacitance and effective-resistance transfer remain ongoing work.

Thank you for the detailed response and additional experiments. However, most of the concerns remain. First, the difference between the original Analog R2R^2 (0.4245) and the new diagnostic result (−0.1440) needs discussion. Second, the new ablations do not show a clear benefit from memory or alignment. Finally, the advantage of domain-conditioned positional encoding over shared encoders under LODO is still missing. I would like to retain my score.

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

This paper addresses the clear problem of ground capacitance prediction across circuit domains, identified as fundamentally important because it directly dictates loading, delay, and switching energy across all circuit types. Authors propose a transformer-based model working on graph-encoded circuit information, evaluating its performance with respect to other SOTA baselines. Results suggest that cross-domain parasitic prediction requires models to separate structural circuit design from domain-specific physical context, rather than just improving raw accuracy.

Critical strengths
  • Relevant problem addressed with a careful protocol - Cold-start parasitic prediction for a new circuit domain is a realistic need. The RQs are clearly separated and evaluated, making the study a useful reference for cross-domain parasitic estimation
  • Broad comparison with consistent gains - Fourteen baselines from four method families are trained under the same protocol. The ablations on all three targets suggest that each component contributes.
Critical weaknesses
  • No data consistency among different circuit domains - Appendix D.1 shows that each domain uses a different technology, so the domain shift mixes circuit type and process technology, and the authors admit it can’t be separated. The study shows transfer across combined circuit and technology shifts, not circuit representation alone.

  • No account for computational overhead - The paper reports no training/inference time, parameter count, or memory footprint for XRC-Former versus the baselines. The appendix also notes that end-to-end preprocessing cost and throughput on full designs were not measured.

  • limited novelty - The method combines known parts as positional encodings, a memory bank, and MMD alignment so the contribution is mainly the application and its evaluation.

  • Possible inconsistencies and typos in the main text - There are typos, like "propsoed" at line 113; one baseline is named inconsistently (CirGPS vs CircuitGPS); the loss is called L_GC in Eq. 2 and L_C in Eq. 11; in Table 1, MLParest and DLPL-Cap have identical standard deviations in all three columns, while their means differ, looking like a copy error; Appendix G says the data are proprietary, but Sec. 5.1 says they come from ParasGB.

Questions for authors
  • Authors are suggested to add requested experiments and details (see weaknesses) to improve results relevance and clarification on which data actually impact cross-domain predictions.

  • The paper requires general adjustments for consistency and errors fixing, this can improve reliability on final evaluation

1 comment

Thank you for the comments on motivation, novelty, computational cost, and reproducibility. We clarify the contribution and address the identified issues below.

1. PDK variation and data scarcity

Circuit datasets are scarce across technologies and design families. Matched implementations and parasitic-extraction labels under multiple PDKs require substantial effort. We therefore study reuse of available source data when a new circuit regime lacks training labels.

Recurring typed topology and local connectivity offer reusable information, while ground capacitance CgC_{\mathrm{g}} also depends on process, geometry, and loading. Our framework combines structural reuse with context-sensitive information, without assuming process-independent capacitance. We agree that the benchmark involves coupled circuit-family and process shifts. We will clarify this scope and the need for a common-PDK control to separate their effects. With adequate target labels, specialist training is a separate setting.

2. Application-level contribution and methodological distinction

Prior work provides valuable predictors for particular settings: MLParest and ParaGraph address custom/analog circuits, ParaFormer studies digital pre-routing RC prediction, and DLPL-Cap targets SRAM capacitance. CircuitGPS and CircuitGCL also explore data-efficient learning and transfer in AMS settings. Our work builds on these advances and focuses on reuse across circuit regimes without target-domain training labels.

The application-level contribution is a cross-domain framework for NET-node GC prediction that addresses scarce, fragmented parasitic labels. It aims to reduce dependence on collecting a new labelled corpus for each regime and provide a practical basis for further data-efficient parasitic research.

The same prototype assignment β\beta retrieves node content M=βZM=\beta Z and contributes a co-assignment term ββ⊤\beta\beta^{\top} to attention, coupling content retrieval with message weighting. MMD regularizes hinvh_{\mathrm{inv}}, while prediction uses hinv+hdomh_{\mathrm{inv}}+h_{\mathrm{dom}}, giving alignment and predictive information retention different routes. These task-specific choices do not guarantee identifiable disentanglement. We will clarify the application contribution and mechanisms in the introduction and related work.

3. Small-scale computational cost

Given limited idle GPU access, we added a small-scale diagnostic on a shared NVIDIA A100-PCIE-40GB (FP32, batch size 1). XRC-Former has 786,574 parameters. On three real 128-node components, GPU-resident inference took 6.90±1.056.90\pm1.05 ms per component, excluding preprocessing and CPU-to-GPU transfer. After 10 warm-ups per graph, we timed 5 blocks of 20 calls with CUDA synchronization; mean/sample SD summarize 15 block means. Peak PyTorch-allocated inference memory was 14.57 MiB, excluding CUDA context and other users. On nine components, a five-epoch training/evaluation workflow took 20.64 s, excluding raw loading; first-pass descriptor preparation took 0.083 s. Shared GPU load affects latency. These measurements establish neither a speed advantage nor full-design throughput. We will state their small-scale scope explicitly.

4. Identical SDs and baseline comparisons

Thank you for identifying the repeated SDs. They resulted from a copy-and-paste error during manuscript table preparation. The corrected DLPL-Cap row in Table 1, in Digital / SRAM / Analog order, is 0.4354±0.07620.4354\pm0.0762 / 0.0123±0.01080.0123\pm0.0108 / −0.3649±0.1890-0.3649\pm0.1890. We will correct these entries and check the remaining table for consistency.

We will distinguish faithful task adaptations from architecture-inspired adapters and document retained modules, replaced heads/objectives, features, budgets, and selection rules. Adapter results should not imply superiority over the original methods in their native tasks.

5. ParasGB, data access, and presentation

ParasGB is a benchmark study for parasitic estimation. We use the same underlying dataset, obtained through a research collaboration. Our study investigates cross-domain reuse of these data under scarce target labels. Access through collaboration does not imply permission to redistribute all underlying circuit or PDK-dependent files. We will specify the dataset version/subset and distinguish publicly accessible artifacts from data requiring permission, clarifying the current public/proprietary wording. We remain committed to further research using open datasets and supporting reproducibility.

We will unify CircuitGPS/CirGPS and GC-loss notation, correct spelling errors, and qualify performance claims by metric and statistic. Table E2 reports lower Digital MAE for CircuitGPS and CircuitGCL. Formal evaluation in this paper concerns GC; preliminary coupling-capacitance and effective-resistance transfer experiments remain incomplete and will be described as ongoing work.

Reviewer_h9AGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper claims that predicting Ground Capacitance, which represents loading to the reference node, can help prioritize design changes and reduce costs in the layout-extraction-verification cycle. It proposes a graph Transformer-based framework, XRC-Former, which includes three major parts that make the framework perform well on the prediction task: 1.Domain-Sensitive Circuit Encoding: The framework incorporates circuit topology, domain-specific layout-derived geometry, and local electrical context into its encoding process. 2. A Shared Structural Memory Mechanism: The framework proposes a shared memory bank where each slot acts as a latent representation learned across source-domain graphs. 3.Invariant-Space Alignment: In the final Transformer layer and after that layer, it uses what the authors call invariant-space alignment to force the whole hidden state to be domain-invariant, the intention behind this is to remove GC-relevant domain context, thereby making the reusable subspace less domain-separable. This encourages the model to align what should transfer. The authors then use experiments to show that XRC-Former improves held-out macro-domain GC prediction and that the invariant subspace learned by XRC-Former becomes less separable across the domains they tested (Digital, Analog, and SRAM). They also conduct an ablation study to show that all three major parts mentioned earlier are necessary for the overall prediction performance.

Critical strengths
  • The paper has practical motivation. The problem it addresses and its generalization could have potential impacts.
  • The architecture is designed for domain generalization. The 3 major innovative parts are clearly presented
  • Leave-on-domain-out evaluation is a meaningful experiment setting
Critical weaknesses
  • One of the major flaws of the paper that cannot be overlooked is the dataset. First, mixing different technology nodes as varied as 180nm BCD for analog, FreePDK 45nm for digital, and TSMC 28nm for SRAM is not typically acceptable or justifiable. Second, unlike digital logic circuits and SRAM circuits, which have stylized layouts, analog layouts can have huge variations and therefore are hard to converge to a single ground capacitance prediction without considering the physical design details. So the premises of predicting a single ground capacitance for an analog circuit may be ill-defined. Finally, the number of nodes and edges specified in Table D.2 suggests that the extracted post-layout netlists of the analog circuits are used in the evaluation, which seems to conflict with the intention of predicting capacitance at an early design stage before layout is performed.
  • Limited evidence for generalization beyond the selected task. Although the Cg prediction task seems reasonable, there is no strong evidence in the paper to support that the prediction of other parasitic parameters can be transferred in the same way. However, the authors claim that the transferable representation and memory mechanisms provide a basis for extending the framework to other parasitics, which is actually an assumption rather than a proven fact.
Questions for authors

Please justify the choice of circuits from different technology nodes and clarify whether post-layout analog netlists are used in the dataset.

1 comment

Thank you for the careful assessment. Our intended contribution is ground-capacitance prediction when labels from a new circuit regime are unavailable. We clarify the motivation, input stage, and scope below.

1. PDK differences and circuit-data scarcity

Our setting is motivated by the limited availability of labelled circuit data across technologies and design families. Constructing matched implementations of the same circuits under multiple PDKs requires substantial design, physical implementation, and parasitic-extraction effort. This makes it valuable to investigate whether knowledge from available circuit corpora can be reused when entering a new regime with no target training labels.

Recurring typed connectivity and local circuit motifs provide a basis for structural reuse, while ground capacitance CgC_{\mathrm{g}} also depends on physical and electrical context. XRC-Former combines structural reuse with context-sensitive encoding and selective alignment. It does not assume that identical topology has identical capacitance under different PDKs.

We agree that the current benchmark involves a combined shift in circuit family and process/flow. Data scarcity motivates this setting, but does not separate these effects. We will make this scope explicit and identify a common-PDK comparison as a necessary control for isolating circuit-family transfer.

2. Analog inputs and the prediction stage

The Analog inputs are netlists from the floorplanning stage. We will state this input stage explicitly in the revised manuscript.

3. Scope beyond ground capacitance

Extending the framework to coupling capacitance CcC_{\mathrm{c}} and effective resistance ReffR_{\mathrm{eff}} is part of our ongoing work. Our preliminary experiments suggest that the cross-domain model can also support transfer learning for these quantities. However, these experiments are not yet complete enough for a systematic evaluation, and we do not present them as established findings of this paper.

We agree that the manuscript should narrow its claims about general parasitic prediction. We will clarify that the present paper evaluates ground capacitance CgC_{\mathrm{g}} and discuss coupling capacitance and effective resistance as promising directions for future work that require further validation.

Comments on the paper

1 comment

Dear Reviewers,

The authors have submitted their rebuttal. Please review it and share your comments. If anything is unclear or requires clarification, please let the authors know as soon as possible. The author–reviewer discussion period ends on 18 November.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.