- 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:
- 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?
- 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?
- 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?
- How do the masking protocols relate to how foundries or PDK teams actually choose characterization points?
- 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.