acceptodds
Under review as a conference paper at ICLR 2027

Chip Capacitance Prediction by a Jointly Routed Transformer

Abstract

Accurate parasitic capacitance estimation is essential for chip design and verification. Lookup-table-based parasitic extraction (LUT-based PEX) reuses precomputed capacitance tables to avoid weeks to months of physical simulations, but sparse coverage limits accuracy for unobserved chip layout geometries. To address this limitation, we propose a jointly routed Transformer for chip capacitance prediction, conditioned on layout geometry, material context, and available capacitance observations. Our key idea is to train a shared Transformer to jointly learn residual corrections to an interpolation estimate and adaptive weights for combining neural and numerical corrections. Notably, the model extrapolates beyond the observed geometry range and enables transfer to an unseen manufacturing process without parameter updates, aided by our structured tokenizer. Moreover, we curate and augment a unified dataset and evaluation suite for capacitance prediction from sparse observations, advancing AI research in LUT-based PEX. Our method outperforms commercial-style multilinear interpolation in all 15 evaluated settings and reduces normalized extrapolation error by 84–86% on public tables. Encouragingly, these gains extend to downstream chip analysis: timing estimates using our predicted capacitances more closely match the full-table reference in all 16 settings across eight fixed layouts.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Chip Capacitance Prediction, Chip Design, AI for EDA, Neural–numerical hybrid modeling, Learned routing, Extrapolation, Parasitic capacitance extraction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Y7HmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper predicts missing entries of capacitance lookup tables used in rule-based parasitic extraction. Each table is one episode. Observed rows and query rows are tokenized together with material-stack tokens, and a small two-layer Transformer (about 105k parameters) reads them. The prediction is an interpolation anchor plus a gated mixture of 11 corrections: zero, two neural residuals (raw and clipped to a local envelope), PCHIP, an aligned-width rule and six fixed RBF-GP predictors. The gate is a per-query, per-component softmax that uses the Transformer state and support statistics. Training uses random, block and boundary-extrapolation masks on 1,713 public rule tables and a FEM reference set. At 30% missingness, the method beats rank-matched commercial multilinear interpolation in all 15 conditions and cuts out-of-support error by 84 to 86% on public tables. It also transfers to ASAP7 without updates, and it improves OpenRCX/STA fidelity in all 16 downstream settings.

Critical strengths

Strength 1: The evaluation protocol is careful. Configuration is chosen on a validation cohort only, the confirmation cohorts are opened once, paired bootstrap intervals are reported, and all methods are scored on the same query population with a shared fallback. For the downstream audit, the ensemble is retrained with all 84 rewritten tables excluded. Code is available through an anonymous link. This level of care is not common in ML-for-EDA papers.

Strength 2: The downstream check is convincing for the stated scope. Writing predicted entries back into OpenRCX tables lowers capacitance error by 1.8 to 13.9x and worst-slack deviation by 2.1 to 56x in all 16 layout settings (Table 3), with fixed routed databases. Source-only transfer to ASAP7 also works without parameter updates (Table 2), which is a practical property for new process nodes.

Critical weaknesses

Weakness 1: The main architectural idea, routing over an 11-branch correction bank, is not clearly needed, and learned baselines are missing. By the authors' own controls (Appendix E.1), full routing improves on a binary gate by only 0.6 to 5.6% on public rules. The ungated residual Transformer is better on in-support block masks, and the binary gate is better on two ASAP7 conditions. A local affine predictor beats the method on out-of-support random masks for both FEM and ASAP7 (Appendix E.2). So most of the gain over interpolation comes from a support-conditioned residual Transformer. Yet that part is never compared with existing in-context regressors (for example attentive neural processes, which the paper cites) or with published neural capacitance models trained on the same tables.

Weakness 2: The benchmark is narrow, and the extrapolation numbers use a weak reference. The public tables vary only one geometric dimension, the FEM reference has 27 evaluation tables, and transfer is tested on one process with 8 tables. The 84 to 86% figure is measured against CMI with endpoint clamping, which by design cannot extrapolate. Against the local affine control, the gap is smaller and not uniform. A physics-based fit (for example a closed-form width/spacing capacitance model fitted to the support) is a natural extrapolation baseline and is not included. It is also unclear whether random and block masks reflect how sparse tables are actually characterized in practice.

Questions for authors

Clarification requests:

  1. How does the method compare with an attentive neural process or a similar set-conditioned regressor trained on the same episodes, and with an existing neural capacitance model (MLP, CNN or GNN) trained per process?
  2. Appendix G reports that pruning the bank at inference changes public-rule NRMSE by at most 0.5%, which suggests that most branches matter little there. Does the same hold on the FEM reference and ASAP7? What does a retrained reduced bank (anchor, neural residual and local affine) give?
  3. How far beyond the support do the out-of-support queries lie, as a fraction of the observed range? How does the error grow with that distance?
  4. How do the masking protocols relate to how foundries or PDK teams actually choose characterization points?
  5. Given prior hybrid rule/neural/field-solver extraction work (for example Abouelyazid et al., 2022), can you state more precisely what the "first neural framework" claim covers?

Suggestions for future work:

The main text is very compressed, and several key results (the matched-budget controls and the local affine comparison) are only in the appendix. Moving them into the main text would give readers a more accurate picture of where the gains come from.

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

This paper addresses the fundamental problem in chip design of accurate parasitic capacitances estimation. It improves the LUT-based PEX sparsity by increasing coverage on unobserved layout geometries with Transfomer predictions combining Transformer architecture, conditional neural processes as context sets, and analytical EDA heuristics. It outperforms recent baselines, bypassing the need for target-domain fine-tuning. Authors also craft a unified dataset across different technologies as foundational benchmark.

Critical strengths
  • ML for LUT table completion and extrapolation is a promising approach not yet explored in the literature
  • Solid comparison with traditional SOTA baselines
Critical weaknesses
  • Real motivation costs not addressed - The real gap addressed in the abstract is that the motivation is cost (“weeks to months” of simulation), yet the paper never quantifies the trade-off between characterization cost saved and accuracy lost. The sparsity introduced by random masks is also synthetic, not a real budget-constrained characterization scenario.

  • Limited methodological novelty, and the novel part adds little - The overall method shows a massive and impressive improvement over the industry standard and authors claim a jointly routed transformer as main contribution. However, Appendix E.1 shows that the core architectural novelty, the branch component-wise routing mechanism, offers only marginal gains over a simpler architecture.

  • Limited scope relative to the stated problem - The introduction contrasts field-solver extraction (accurate but expensive) with lookup-table extraction (fast but less accurate). The method addresses only one error source of the second: interpolation over missing entries in sparsely characterized tables. Its best case is therefore recovering the full table. It does not reduce the systematic gap between table-based and solver-based extraction caused by pattern approximation.

  • Unclear and too verbose sections, dense explanation - Main ideas on the workflow are lost during the read. For example, Figure 1 and Figure 2 are too detailed and their steps and stages are not consistently developed in the sections of the paper.

Questions for authors
  • In Appendix E.1, full routing beats the simpler controls only marginally on public rules, and the binary gate wins two ASAP7 transfer conditions. Gains are clearer only on the FEM reference. Given the inference cost, is the 11-branch bank necessary? When would you recommend the simpler residual model instead?

  • How much solver or characterization time does a given missingness rate save in a realistic flow? Can you show accuracy as a function of the number of characterized points against simply characterizing more points?

Reviewer_u6CsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a jointly routed Transformer framework for predicting missing parasitic capacitance values in sparse lookup tables. The method combines Transformer-based residual prediction with conventional numerical estimators through a learned routing mechanism. Experiments on public extraction-rule tables, finite-element reference data, and cross-process transfer show lower NRMSE than conventional interpolation baselines, while downstream evaluations demonstrate better agreement with full-table capacitance extraction and timing estimates.

Critical strengths
  1. The paper evaluates the proposed approach across multiple technology settings, including public extraction-rule tables, finite-element reference data, and cross-process transfer.

  2. The evaluation includes conventional numerical baselines, simpler learned-model ablations, and downstream parasitic extraction and static timing analysis on fixed routed layouts, providing evidence beyond capacitance prediction error alone.

Critical weaknesses
  1. The proposed framework primarily combines established techniques, including Transformer-based prediction, residual learning, numerical interpolation, and learned routing. The paper does not clearly establish its methodological novelty or provide sufficient justification for its architectural choices, particularly the use of a Transformer and an elaborate 11-branch routing mechanism for relatively low-dimensional tabular data. The rationale for this design over simpler alternatives is not adequately developed.

  2. The evaluation relies heavily on artificially masked lookup tables, without clearly establishing how these settings reflect practical characterization workflows or generalize to more complex geometric configurations. Additionally, the manuscript devotes excessive space to mathematical formulations of standard operations rather than deeper methodological insights and design justifications. Figures 1, 2, and 3 are difficult to read, and numerous formatting and spacing issues further detract from the presentation.

Questions for authors
  1. What motivates the choice of a Transformer for this task, and which properties of the capacitance data make it preferable to simpler architectures? Could the authors provide more insight into the design of the geometry and observation tokenization?

  2. What is the rationale for including all 11 correction branches? Given the pruning results reported in the appendix, could the authors clarify which branches are essential and whether a substantially simpler routing mechanism could achieve comparable performance?

  3. How do the proposed method and its methodological contributions differ from existing neural, residual, and hybrid approaches to capacitance prediction? What is the main new machine learning insight beyond combining these established components?

  4. How representative are the random, block, and boundary-extrapolation masks of missing observations in practical capacitance-table characterization? What realistic characterization scenarios motivate the selected missingness rates?

  5. How should the reported NRMSE improvements be interpreted in terms of practical extraction accuracy? Could the authors provide additional insight into the distribution of prediction errors, particularly for difficult geometries or capacitance components where large individual errors may matter?

The manuscript would also benefit from substantial presentation improvements. In particular, Figures 1, 2, and 3 should be made more readable, and the numerous spacing issues should be corrected. Unnecessary mathematical exposition of standard operations could be removed, with the resulting space used to bring important analyses and discussions from the appendix into the main paper, particularly those that clarify the methodological contributions and justify the architectural design choices.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.