- 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, . 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 . 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 and 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 response and additional experiments. However, most of the concerns remain. First, the difference between the original Analog (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.
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(6log10(max(C,10−30)1020),0,1), TD(C)=clip(6log10(max(C,10−30)106),0,1).
Analog/Digital inputs use F/pF, respectively. These are domain-dependent label transforms, with a 1/3 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
Winv and Wdom 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.
The distinction lies in the training objective: MMD regularizes hinv, 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:
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 R2 gain is 0.2405±0.0219 (population SD), with higher R2 in all five seeds. This comparison concerns the GeoPE/PhysPE paths jointly.
Separately, five-seed real-three-domain raw-data runs with K=64 provide source-training hard-assignment statistics at source-validation-selected epochs. For the Analog-target fold, occupied slots are 54.8±1.1, normalized entropy H/ln64=.864±.017, and top-slot share (8.01±1.95)% (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 R2 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.