acceptodds
Under review as a conference paper at ICLR 2027

ArrivalFM-OTA: A Large-Scale Benchmark for Physical Representation Learning in Analog Circuit Design

Abstract

Surrogate modeling in analog integrated circuit design is critically obstructed by stiff nonlinear differential-algebraic equations, discrete device sizing, and a scarcity of open, standardized benchmarks and datasets. To bridge this gap, we introduce ArrivalFM-OTA, a forward modeling (FM) benchmark and dataset for physical representation learning in analog circuit design. ArrivalFM-OTA comprises over 2 million industrial-grade Cadence Spectre simulation datapoints across 28 operational transconductance amplifier (OTA) topologies and 10 FinFET technology nodes from 7nm to 20nm, evaluated under continuous environmental sweeps. On top of this data, ArrivalFM-OTA establishes a surrogate modeling challenge across 4 standardized distribution splits and 68 joint prediction targets spanning validity flags, global performance metrics, and internal transistor operating states. Evaluating representative neural baselines across extensive multi-seed initializations, we propose DAWN (DC Operating Point Analog Waypoint Nodalization), a nodal-conditioned residual framework that improves global metric prediction across all backbones by up to +0.055 R², with our leading model reaching R² ≈ 0.793 in-distribution and R² ≈ 0.764 on an unseen technology node. However, performance drops to R² ≈ 0.281 on unseen topologies, revealing that while neural models achieve competitive interpolation across continuous physical parameters, structural extrapolation across circuit graphs remains a fundamental open challenge for physical representation learning. Finally, we open-source all data, benchmark protocols, multi-seed baseline profiles, and our simulation orchestration pipeline to provide a standardized foundation for transistor-level machine learning and circuit design workflows.

Primary area:
datasets and benchmarks
Keywords:
Benchmarks and Datasets, Electronic Design Automation, Scientific Machine Learning, Analog Circuit Design, Neural Surrogate Modeling, Out-of-Distribution Generalization, Geometric Deep Learning

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_mU6COpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

ArrivalFM-OTA presents a dataset containing 2,649,573 Spectre simulation records for 28 OTA topologies and 10 FinFET technology models for prediction purposes, where 2,059,328 simulation records are grouped to create a benchmark with 68 global and device-level predictions. The paper also presents a residual augmentation called DAWN, which helps in device operating point and circuit performance prediction. Experiments across four neural backbones and 10 initialization seeds demonstrate improvements on known topologies and a held-out technology model.

Critical strengths

S1. The dataset offers abundant supervision for circuit representation learning.
The fact that there is information on circuit performance, internal operating state of transistors, and failure related target under different environmental settings makes a valuable contribution. It helps in understanding the study on the relationship between local behavior and global circuit performance, which is hard to conduct using datasets that provide only aggregate metrics.

Critical weaknesses

W1. The contribution of the DAWN methodological insight with respect to multitask residual augmentation is not clear.
The baselines already receive device-level supervision, while DAWN incorporates increased capacity and modifies how the global context is incorporated into the representations for the task of circuit prediction. The results presented do not separate the impact of these modifications from the introduced physical-conditioning process. Furthermore, the circuit-level predictor utilizes shared latent representations instead of explicit prediction of the operating point, which is not obvious because of the terminology used in the paper ("waypoint" and "physics-conditioned"). This is worsened by the dense formulation, where numerous equations and implementation details do not lead to an understanding of the physical dependency captured and how and when it is verified.

W2. The benchmark’s physical validity and the interpretation of its evaluation metrics require stronger justification.
There are still several points related to data and assessment which require further explanation. Section 3.2 does not explain how the entire data set is used to form the benchmark subset. Appendix E.2 introduces physical failures, numerical failures, and intentionally skipped analyses, while several classification targets in Table 23 use the availability of metrics as criteria for classification. Thus, the average classification accuracy alone does not mean reliable detection of physically invalid circuits. Moreover, Table 22 states the maximum static power of 9.76×1089.76\times10^8 W and mean power of 1244.12 W, but the median power is 3.71×10−43.71\times10^{-4} W, requiring a clear physical analysis of the benchmark selection criteria. These points do not mean the inappropriate nature of the PTM-based data, but they cast doubts about the goals of the benchmark.

Questions for authors

Q1. Clarification. Methodological evidence.
Given that the baseline models already receive device-level supervision, what evidence attributes DAWN’s gains specifically to learning physical intermediate states? Does the advantage persist against a parameter-matched residual baseline with equivalent access to global context, and how does removing device-level supervision affect that comparison?

Q2. Clarification. Physical pathway.
Global predictions in Equations 6 and 7 make use of a latent residual that is common to all of them rather than predictions of operating point values directly. What is the characteristic of this approach that can be used to explain why it may be called physical intermediate-state cascade and what proof exists for the usage of learned device physics in global predictions? How can we justify the statement about zero initialization “completely preventing auxiliary gradient interference”?

Q3. Clarification. Validity of benchmarks.
Why is there a very high power in Table 22, and are the associated data kept in the training and validation sets? Please specify how the 2,059,328 benchmarking data samples are selected, and if any improvement in the results can be observed by removing artifacts.

Q4. Clarification. Failure prediction accuracy.
How does Table 23’s labeling distinguish intentionally skipped analyses from physical failures, particularly when periodic analysis is skipped because unity-gain bandwidth is at most 1 kHz? What are the class frequencies and precision and recall of failure classes after making such a differentiation? How should classification accuracy be evaluated without making such a differentiation?

5 comments
AuthorsOpenReview ↗

Response To Reviewer mU6C (Introduction, W1, Q1, Q2)

We would like to thank Reviewer mU6C for their rigorous and insightful critique of our submission. We are grateful for your recognition of dataset scale and the value of providing dense multi-task supervision spanning circuit performance, internal device operating states, and characterization flags under environmental sweeps. We are especially appreciative of your audit regarding classification criteria and numerical artifacts in our benchmark and dataset, both of which have been clarified in revised versions of Section 1 (Pages 1-3), Appendix F (Pages 45-48), Appendix G (49-51), and a newly instantiated Appendix I (Pages 57-60).

W1 & Q1: We thank Reviewer mU6C for their raising this point. We agree with your assessment: because baseline models already receive device-level supervision, our original manuscript blurred the distinction between targeted bottleneck capacity and multi-task guidance. In direct response to your question: we do not have an existing baseline that isolates the +148,224 parameter bottleneck (+0.15M weights) from auxiliary device supervision, nor an ablation removing device supervision from the bottleneck.

Consequently, we do not claim that DAWN's gains stem from an isolated physical mechanism independent of capacity. In standard baselines, device-level supervision and global specification predictions are decoupled into independent heads operating on the final backbone representation hv(L)\mathbf{h}_v^{(L)}. DAWN modifies this by injecting an auxiliary-guided latent residual stream into the shared representation. Its empirical lift reflects the combined benefit of this targeted bottleneck capacity and multi-task guidance. Two empirical findings clarify this role.

First, adding DAWN's +148k parameter bottleneck and auxiliary supervision to our nonrelational MLP baseline yields only +0.0150 R², contrasting with +0.0437 R² across a parameter-matched, graph-based GCN backbone. This demonstrates that while added capacity yields a minor baseline lift, DAWN’s empirical gain is strongly amplified by relational component connectivity.

Secondly, R² gains are not uniform across targets as would be expected from generic capacity expansion. In our revised manuscript, we have added Table 32 (Appendix I, Page 59) establishing a full per-metric breakdown of R² gains across our DAWN ablations. Please see our posted comment Response To Reviewer 1ZSL (W3, Q1 Part I) for exact numerical details.

W1 & Q2: We thank the reviewer for highlighting these distinctions. Your observations are correct: the circuit-level predictor utilizes a shared continuous latent residual rather than explicit numerical operating points, and zero-initialization does not prevent auxiliary gradient flow during training. We have updated Section 4.2 (Page 7) to state this explicitly: zero-initialization "preserves the backbone’s initial forward mapping. Although auxiliary heads still backpropagate gradients through the identity path, this zero-initialization prevents uninitialized projections from destabilizing early training across all four evaluated backbones".

We also concede that calling DAWN a "physical intermediate-state cascade" and utilizing vocabulary such as "analog waypoints" or "physics-conditioned" was misleading. In reality, global readouts make use of a continuous latent residual vector Δv∈R512\boldsymbol{\Delta}_v \in \mathbb{R}^{512} that is shared across heads. In the revised manuscript, we have systematically removed those terms describing DAWN across the Title, Abstract, Section 1, Section 4.2, Section 5 (Pages 8-10), and Appendix H.2.5 (Page 56), and concede that there is no causal proof that the global readout head explicitly uses learned device physics beyond standard multi-task latent representation sharing. We now describe DAWN objectively as an additive multi-task residual regularizer rather than an autonomous physical solver, and instead focus our analysis on the empirical benefit of DAWN as a whole and the narrowed zero-initialization guarentee. We have also revised our description of DAWN regarding DC operating points (DCOP) as a DCOP-Aware Weighted Network, from its original description of DCOP Analog Waypoint Nodalization, to more accurately reflect technical reality, and added Figure 21 (Appendix I, Page 56) as an informative visual reference of the DAWN framework.

AuthorsOpenReview ↗

Response To Reviewer mU6C (W2, Q3)

W2 & Q3: We thank Reviewer mU6C for these vital questions regarding dataset partitioning and dataset convergence artifacts.

In Section 1 (Pages 1-3) and Section 3.2 (Page 5), we clarify that the standardized ArrivalFM-OTA benchmark consists of 2,059,328 simulation records, gathered from running data collection for each of our 28 OTA topologies strictly following Algorithm 1 (Section 3.1, Page 4). The remaining 590,245 datapoints in our dataset represent 100% physically valid Spectre simulations from early exploratory sizing sweeps, which are released alongside our open-source Contact synthesis framework as an unpartitioned archive for open self-supervised pretraining. The 2,059,328 benchmark datapoints are then partitioned into Training, Validation, and 4 standardized Test splits, as defined in Table 2 (Section 3.2, Page 5).

Power consumption across ArrivalFM-OTA is physically grounded and tightly clustered around a median of P = 0.371 mW with an interquartile range of IQR = 2.16 mW, where over 99.5% of evaluated designs draw less than 4.74 mW. The divergence in the arithmetic mean P = 1244.12 W and maximum value 9.76×1089.76 \times 10^8 W arises from a numerical Spectre solver artifact affecting exactly 309 of 2,059,328 total benchmark samples, occurring exclusively within unfeasible designs of a single topology, OTA_FD03. DC open loop gains for these 309 samples range from -134.09 dB to -11.15 dB, drawing from 50.06 mW to 976.21 MW of power. We hypothesize that severe voltage headroom starvation in OTA_FD03's 5-transistor stack under unclamped ideal common-mode feedback (CMFB) caused simulator divergence on the auxiliary supply current variable V0:p. Furthermore, physical transistor currents in these exact simulations remained below 3.71 mA. These design points are intentionally retained in the benchmark as negative training samples for feasibility classifiers and component-level metric regressors. We have included this explanation of these anomalies in the description of Table 23 (Appendix F, Page 48).

Additionally, in direct response to your question regarding whether removing artifacts improves performance: because neural surrogate models optimize Z-score normalized loss objectives and are evaluated via rank-based and median statistics, these 309 non-convergent points do not destabilize model training or corrupt median metrics. As reported in Table 6 (Section 5, Page 10), our DAWN GCN achieves a static power physical MedAE of 103.37 ± 1.83 µW (R² = 0.9562 ± 0.0041) across all 161,160 in-distribution test samples. However, unweighted arithmetic MAE is inherently vulnerable to extreme numerical outliers. As explicitly detailed in Footnote † of Table 6, when static power physical MAE is evaluated across the remaining 19 in-distribution topologies excluding OTA_FD03, DAWN GCN achieves an MAE of 879.24 ± 23.90 µW, directly matching physical sub-milliwatt circuit scales. In sharp contrast, if OTA_FD03 is included, the arithmetic physical MAE across the test set explodes to 204.51 W, driven entirely by those divergent simulator artifacts. Removing this single artifact topology reduces the physical MAE by more than five orders of magnitude (from 204.51 W down to 879.24 µW), verifying that model predictions outside this simulator singularity remain physically faithful.

AuthorsOpenReview ↗

Response To Reviewer mU6C (W2, Q4 Part I)

W2 & Q4: We thank Reviewer mU6C for their critique regarding the semantic distinction between physical circuit failures, simulator numerical divergences, and intentionally bypassed analyses. We fully agree that evaluating surrogate fidelity solely via aggregate macro accuracy could obscure failure-mode detection if criteria were conflated.

We have revised our description of classifier heads in Table 24 (Appendix G, Page 50) to provide insight into our reasoning for introducing classifier heads as a universal diagnostic for defined NaN failures in Appendix D.2 (Pages 31-34) and Appendix E.2 (Pages 36-37). Downstream analyses such as large-signal dynamic swing (swing_stab), DC offset (offset_stab), transient settling envelopes (transient_sr, transient_sf, transient_lr, transient_lf), and periodic steady-state analysis (periodic_stab) are trained to classify an OTA design as invalid (0) if a candidate fails an upstream AC/DC gatekeeper as described in Appendix E.2 or fails to return a valid result given target performance metric definitions in Appendix E.3 (Pages 37-44). Specifically, for periodic steady-state analysis, we require Cadence Spectre's harmonic shooting algorithm to have a stabilization time of tstab=100/fugbwt_{\text{stab}} = 100 / f_{\text{ugbw}}. When unity-gain bandwidth collapses below 1 kHz, the required stabilization window extends to tstab≥100 mst_{\text{stab}} \ge 100\text{ ms}, which we observe commonly causes divergence errors and excessive CPU timeouts. Thus, each head acts as an independent, domain-specific feasibility filter for determining whether a candidate sizing yields valid, non-divergent specifications within that respective characterization domain, while complete end-to-end simulation convergence is independently isolated by the validity head, conditioned on sim_status as defined in Table 21 (Appendix E.2, Page 37). Notably, sim_status retains nominal validity even when simulation stages are declared as invalid under the fugbwf_{\text{ugbw}} ≤ 1 kHz constraint.

To directly address your question regarding class frequencies and whether accuracy is inflated without differentiating failure classes, we have added Table 30 (Appendix I, Page 57), which reports decoupled performance across all 10 individual classification heads for the DAWN GCN baseline as mean ± standard deviation across 10 random initializations:

Classifier HeadGround Truth Positive RatePredicted Positive RateClassification AccuracyPrecisionRecallF1\text{F}_1-Score
validity63.24%64.67% ± 0.34%94.79% ± 0.06%0.9488 ± 0.00270.9701 ± 0.00240.9593 ± 0.0004
nominal_stab67.55%68.71% ± 0.27%96.16% ± 0.05%0.9636 ± 0.00210.9801 ± 0.00170.9718 ± 0.0004
threshold_stab95.24%95.38% ± 0.15%99.12% ± 0.02%0.9946 ± 0.00080.9961 ± 0.00080.9954 ± 0.0001
swing_stab61.29%62.37% ± 0.35%95.44% ± 0.05%0.9547 ± 0.00280.9717 ± 0.00270.9631 ± 0.0004
offset_stab61.29%62.38% ± 0.35%95.44% ± 0.05%0.9547 ± 0.00280.9717 ± 0.00270.9631 ± 0.0004
periodic_stab60.63%61.94% ± 0.38%94.69% ± 0.08%0.9466 ± 0.00310.9669 ± 0.00290.9566 ± 0.0006
transient_sr60.04%61.60% ± 0.39%93.64% ± 0.06%0.9357 ± 0.00290.9600 ± 0.00310.9477 ± 0.0005
transient_sf60.63%62.28% ± 0.37%93.55% ± 0.06%0.9350 ± 0.00260.9604 ± 0.00310.9475 ± 0.0005
transient_lr50.45%51.61% ± 0.39%92.57% ± 0.06%0.9168 ± 0.00300.9379 ± 0.00420.9272 ± 0.0008
transient_lf46.74%47.99% ± 0.39%92.20% ± 0.08%0.9057 ± 0.00320.9300 ± 0.00450.9177 ± 0.0010
Reviewer_J5n7OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

ArrivalFM-OTA introduces a dataset and standardized forward-modeling benchmark for operational transconductance amplifiers. The benchmark contains 2,059,328 examples across 28 topologies and ten FinFET predictive technology-model variants, drawn from a larger 2,649,573-simulation collection. Its 68 targets combine global performance, device operating points and diagnostic classifications. Four test regimes separate in-distribution, topology, technology and combined holdouts. Four backbone families and their DAWN variants are evaluated across ten initializations. DAWN adds a shared residual bottleneck before local and global readouts. Results show consistent in-distribution and technology-holdout gains, with markedly less reliable transfer to unseen circuit structures.

Critical strengths
  1. The benchmark has substantial independent AI/ML value. Dense device-level targets, circuit-level specifications, diagnostic outcomes and explicit domain-shift splits make this a useful resource for studying structured physical representations. Its contribution is more specific than dataset size: ZeroSim already supplies millions of examples across more topologies, whereas Arrival adds a different combination of internal states, technology conditions and standardized tasks. The dense device-level labels and standardized evaluation protocols are useful assets for studying representation learning within the OTA domain.
  2. The evaluation characterizes both successes and limitations. Ten-seed results on the fixed splits and per-topology breakdowns offer useful evidence about stability and structural generalization. DAWN GCN raises ID global R² from 0.7492 to 0.7929 and technology-holdout R² from 0.7160 to 0.7639, while topology-holdout R² does not improve (Table 4); the main text acknowledges mixed OOD outcomes. The anonymous release includes source, saved configurations and an aggregate results file consistent with the reported table. These are meaningful reproducibility assets.
Critical weaknesses
  1. The resource and model documentation need reconciliation with the released implementation. Table 26 reports 2.15M ParaGraph and 2.29M GT parameters at width 512/depth 4. However, the public ParaGraph uses nine full relation matrices, already requiring 2,359,296 weights before its GRU and readouts. GT uses four TransformerConv layers with four 512-dimensional heads; concat=False averages outputs without removing the separate projections. The saved configurations support these dimensions, and static layer counting gives approximately 5.30M and 18.97M total parameters. This undermines the claim of matched capacity across backbones, although it does not invalidate within-backbone DAWN gains. Label documentation also needs version binding: the public source uses a 20τ transient setting and includes the phase-margin correction absent from the printed equations; its measurement manager converts dB gain to linear gain. Thus, the source avoids the apparent calculation errors, and I do not infer corrupted labels from the manuscript formulas. The remaining issue is establishing which source/configuration generated the released benchmark.
  2. DAWN's physical-mechanism explanation exceeds the available evidence. All baselines already receive local and global supervision. DAWN routes a shared latent residual, rather than explicitly conditioning on predicted numerical operating points, so its gains may reflect extra nonlinear capacity or residual conditioning. A capacity-matched residual control or an existing auxiliary-supervision ablation would distinguish these explanations. Zero initialization correctly preserves the initial forward mapping, but does not completely prevent auxiliary gradient interference: local heads still backpropagate through the identity path into the backbone (Equations 6–7). Related work also needs precise framing: INSIGHT already exploits dependencies between cheaper and more expensive circuit metrics, although it does not provide this device-level benchmark. Finally, the supported generalization finding concerns one held-out PTM variant, 14nm LSTP, and four held-out topologies; it should not be generalized to new foundries or all novel circuit structures.
  3. Despite its scale and multiple OTA topologies, the benchmark remains confined to a single circuit family, leaving cross-family generalization untested and its broader value for physical representation learning in analog circuit design insufficiently established.
Questions for authors
  1. Which source revisions and saved configurations produced the released labels and Table 4? Can an existing manifest or trace reconcile the transient/phase conventions and the actual trainable-parameter counts?
  2. Is there an existing control separating DAWN's residual capacity from the proposed device-state mechanism? If not, can the explanation be narrowed to the demonstrated empirical benefit and the correct initialization guarantee?
4 comments
AuthorsOpenReview ↗

Response To Reviewer J5n7 (introduction, W1, Q1)

We would like to thank Reviewer J5n7 for their exceptionally rigorous and constructive review of our submission. We are grateful for your audit of our released code and configurations, as well as your recognition of our 10-seed benchmarking protocols, comprehensive per-topology breakdowns, and the reproducibility value of our open source release assets. To further reinforce the empirical rigor of the benchmark, we have added 5 new tables and 2 new scaling figures across Section 5 (Pages 8–10) and a newly instantiated Appendix I (Pages 57–60). These additions provide exhaustive unaggregated evaluations, breaking down internal component-level device operating point metrics (Table 31, Appendix I, Page 58), physical errors in real circuit units (Table 6, Section 5, Page 10), decoupled classification diagnostics across all 10 characterization heads (Table 30, Appendix I, Page 57), cross-quadrant out-of-distribution physical deltas (Table 33, Appendix I, Page 60), and empirical topological complexity scaling correlations with global metric prediction (Figure 3, Section 5, Page 10), (Figure 22, Appendix I, Page 58).

W1: We thank the reviewer for their thorough audit. Your deductions are correct: our released code implementation has always used the correct physical calculations and model definitions, while the discrepancies existed solely in our manuscript text.

First, your static layer counts are exact. Table 26 (Appendix H.1, Page 52) previously reported theoretical estimates based on single-matrix abstractions rather than the instantiated PyG modules. We have audited all model state dictionaries and updated Table 26, Section 4.2 (Page 7), and Section 5 (Page 8) to conform to our updated scope and explanations:

Architecture BackboneBaseline ParamsBaseline R²DAWN ParamsDAWN R²Bottleneck AddedMacro-Averaged ΔR²
MLP (No graph edges E\mathcal{E})2,160,198 (2.16M)0.5600 ± 0.00762,308,422 (2.31M)0.5750 ± 0.0055+148,224 (+0.15M)+0.0150
GCN2,165,318 (2.17M)0.7492 ± 0.00392,313,542 (2.32M)0.7929 ± 0.0028+148,224 (+0.15M)+0.0437
ParaGraph5,304,390 (5.30M)0.6975 ± 0.00415,452,614 (5.45M)0.7301 ± 0.0032+148,224 (+0.15M)+0.0326
GT18,973,254 (18.97M)0.7116 ± 0.003719,121,478 (19.12M)0.7499 ± 0.0031+148,224 (+0.15M)+0.0383

As a result, we have removed all claims that capacity was matched across backbone families due to the architectures representing distinct capacity regimes (2.16M to 18.97M). However, within each backbone family, the DAWN module is strictly controlled: every DAWN variant adds the exact same 148,224 parameters via an identical 512 → 128 → 128 → 512 bottleneck.

W1 & Q1: All simulation records in our dataset including non-benchmark datapoints and Table 4 baseline scores were generated by the frozen implementation in spec_functions.py and the training configurations in training/{mlp,gcn,paragraph,gt}.py (width=512, depth=4) in the public repository, which have remained frozen since benchmark data collection began. The text in Appendix D.9 (Page 34) & Appendix E.3.4 (Page 38) was drafted prior to data collection where measurement routines were updated to prevent edge-case artifacts, but the manuscript formulas were regrettably not updated before paper submission. In the revised manuscript, we have updated the following equations to match spec_functions.py:

  • Phase Margin Branch Unwrapping (Equation 40, Appendix E.3.4, Page 38):

    PM=((∠T(fugbw)−Θbranch+360∘) mod 360∘)−180∘\mathrm{PM} = \Big( (\angle T(f_\mathrm{ugbw}) - \Theta_\mathrm{branch} + 360^{\circ}) \bmod 360^{\circ} \Big) - 180^{\circ}

    where Θbranch\Theta_{\mathrm{branch}} resolves multi-sheet phase wrapping at high frequencies.

  • Transient Simulation Time Horizons (Equations 31, 32, & 33, Appendix D.9, Page 34):

    tran tran stop=(1.2 * stop time)\texttt{tran tran stop=(1.2 * stop time)}

    τ=12π⋅fugbw,stop time=20τ=202π⋅fugbw\tau = \frac{1}{2\pi \cdot f_\mathrm{ugbw}}, \qquad \texttt{stop time} = 20\tau = \frac{20}{2\pi \cdot f_\mathrm{ugbw}}

    step time=max⁡(0.001⋅stop time,1 ps)\texttt{step time} = \max(0.001 \cdot \texttt{stop time}, 1\text{ ps})

    tdelay=0.2⋅stop time=4τt_\text{delay} = 0.2 \cdot \texttt{stop time} = 4\tau

  • DC Open Loop Gain Conversion (Appendix E.3.2, Page 37): Clarified that the measurement manager converts open-loop gain to linear units prior to characteristic impedance extraction.

AuthorsOpenReview ↗

Response To Reviewer J5n7 (W2, W3, Q2)

W2 & Q2: We fully agree with the reviewer's assessment: our original text overstated DAWN as an autonomous physical mechanism and blurred the boundary between targeted residual capacity and auxiliary guidance. In response to your specific question: we do not have an existing control that isolates DAWN's +148,224 parameter bottleneck from its auxiliary device-level supervision. Following your recommendation, we have explicitly narrowed our claims in Section 4.2 (Page 7), Section 5, Section 6 (Pages 9-10), & Appendix H.2.5 (Page 56) to anchor our claims strictly to the demonstrated empirical benefit, the relational negative control, and the correct initialization guarantee.

We concede that DAWN does not condition on explicit numerical operating points, but rather routes a continuous latent residual across heads. We have removed all terminology implying an autonomous physical solver such as "physics-informed neural networks" and "analog waypoints". In Section 4.2 and Appendix H.2.5, DAWN is now formalized strictly as a low-rank residual bottleneck. We have also revised our description of DAWN regarding DC operating points (DCOP) as a DCOP-Aware Weighted Network, from its original description of DCOP Analog Waypoint Nodalization, to more accurately reflect technical reality, and added Figure 21 (Appendix I, Page 56) as an informative visual reference of the DAWN framework.

As you correctly noted, zero-initialization does not prevent auxiliary gradient interference: local heads still backpropagate through the identity path into the backbone. We have updated Section 4.2 to state this explicitly: zero-initialization "preserves the backbone’s initial forward mapping. Although auxiliary heads still backpropagate gradients through the identity path, this zero-initialization prevents uninitialized projections from destabilizing early training across all four evaluated backbones".

Because DAWN introduces extra capacity alongside auxiliary supervision into the residual stream, its empirical lift reflects the combined benefit of this targeted capacity and multi-task guidance. Specification-level regression deltas in a newly added Table 32 (Appendix I, Page 59) reveals that this lift is heavily concentrated in dynamic stability and transient metrics rather than distributed uniformly across targets; please see our posted comment Response To Reviewer 1ZSL (W3, Q1 Part I) for exact numerical details.

Additionally, we have updated Section 1 (Page 1) of our paper to accurately contextualize INSIGHT (Poddar et al., 2025), framing its contribution as utilizing an autoregressive Transformer to "predict compute-intensive transient specifications conditioned on cheaper small-signal metrics". Furthermore, we have expanded our prior work comparison in Table 1 (Section 2, Page 2) to include LOADBench (Leibl et al., 2026), concurrent work at the time of submission. LOADBench supplies massive sample counts (100M+) on a legacy SkyWater 130nm CMOS node via Ngspice, further reinforcing your assessment that our contribution lies in dense device-level representations and standardized evaluation protocols rather than raw sample volume alone.

W2 & W3: We agree with the reviewer that our results must not be generalized to arbitrary process nodes or analog circuits as a whole. We have systematically overhauled the manuscript to scope all contributions and conclusions strictly to the evaluated domain:

  1. Title: Explicitly scoped to "ArrivalFM-OTA: A Large-Scale Benchmark for Physical Representation Learning in Operational Transconductance Amplifiers".
  2. Abstract & Section 1: Reframed scope of contribution to a "physical representation learning benchmark comprising over 2 million industrial-grade Cadence Spectre simulation datapoints across 28 operational transconductance amplifier topologies and 10 FinFET technology models from 7nm to 20nm, evaluated under continuous environmental sweeps", and tightened claims of unseen topological transfer to within the scope of our benchmark itself.
  3. Sections 5 & 6, Appendix I: Reanchored to focus on core benchmark diagnostics across the 28 OTA topologies, explicitly discussing why cross-topology transfer is bounded within this domain.

Establishing an open-source benchmark spanning 28 distinct OTA topologies across 10 technology models provides the foundational infrastructure and standardized evaluation splits needed to rigorously study physical representation learning in OTAs. While our current data and evaluation infrastructure is specialized for OTAs only, we view extending this device-level physical representation paradigm connecting local internal operating states to global circuit performance to other analog circuit classes as vital future work.

Reviewer_J5n7OpenReview ↗

Human-written assessment of the paper

I believe the most valuable contribution of this paper is the dataset itself. It establishes a relatively large-scale OTA benchmark that includes not only global performance metrics but also transistor-level DC operating points and simulation diagnostics. Such a dataset provides a useful resource for studying the relationship between internal transistor states and overall circuit performance.

However, I remain cautious about the methodological contribution of DAWN.

After carefully examining its network architecture, I believe there is a gap between the physical conditioning mechanism emphasized by the authors and what is actually implemented. DAWN appears to enhance the model through a shared latent residual representation, rather than explicitly using predicted numerical DC operating points to guide global performance prediction. Considering that the original baselines already receive both local and global supervision, more targeted control experiments are needed to demonstrate that the observed improvements genuinely arise from the claimed physical mechanism.

In addition, I examined the publicly released code and checked the reported model parameter counts. The actual parameter counts of ParaGraph and GT are inconsistent with those reported in the paper, which affects the fairness of comparisons across different backbone architectures. However, I do not believe this discrepancy invalidates the DAWN comparisons within each individual backbone.

Regarding dataset generation, I also identified inconsistencies between some equations in the manuscript and the actual implementation. The authors should clearly specify the source revision and configuration that were used to generate the benchmark.

Overall, I recognize the independent value of the dataset, but believe that the explanation of DAWN's underlying mechanism, model parameter reporting, and implementation reproducibility still have important limitations. I therefore give this paper a Weak Reject recommendation.

Reviewer_J5n7OpenReview ↗

Final assessment after author response

Thank you for the detailed responses and revisions. I acknowledge the corrected parameter counts, clarified measurement descriptions, and narrower characterization of DAWN. However, three concerns remain.

1. DAWN’s distinct methodological contribution remains insufficiently established. All baselines already receive DCOP supervision, while DAWN jointly adds residual capacity and pre-pooling context conditioning. No matched control or device-supervision ablation separates these effects. Different gains across MLP and GCN, or larger gains on selected dynamic targets, do not provide this separation. The revised framing is more accurate, but the evidence establishes the benefit of the combined modification without identifying what makes it specifically effective.

2. The treatment of known regression-label artifacts remains unresolved. The authors identify 309 power values as numerical artifacts, yet the documented finite-value mask in Eq. (9) does not exclude finite erroneous targets. Retaining failed designs for classification does not establish the validity of their power labels. Log normalization and Huber loss do not resolve this issue, and excluding OTA_FD03 from the reported physical MAE does not replace a target-specific correction and re-evaluation.

3. The device-level interpretation conflicts with the reported results. As reported in Table 5, ParaGraph’s operating-region F1 decreases from 0.6798 to 0.0068 with DAWN, whereas Section 5 states that this quantity improves. This discrepancy needs reconciliation before the claimed consistent improvement in device-state prediction is supported.

I recognize the dataset’s value and the improved transparency. Nevertheless, the remaining concerns about methodological significance, regression-target validity, and consistency of the empirical conclusions prevent me from supporting acceptance. I therefore maintain my Weak Reject recommendation.

Reviewer_1ZSLOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This work propose ArrivalFM-OTA which is a forward modeling benchmark and dataset, aiming to represent physical learning in analog circuit design. They include 28 OTAs under 10 different tech nodes and they use Cadence to simulate all of them. They also proposed DAWN (DC Operating Point Analog Waypoint Nodalization), which is a nodal-conditioned residual cascade that routes intermediate operating point representations through zero-initialized projections, and conduct an exhaustive evaluation of 8 model architectures spanning 4 distribution splits and 10 random initializations each. The OTA are separated into 10 easy single-ended output ones, 3 fully differential output ones, 4 medium single ended ones, 8 fully differential output ones and 3 hard fully differential output OTAs. The benchmark contains 43 performance metrics and device metrics like gm gds. The sampling points are sampled from RQMC Sobol XGBoost Surrogate sampling to better represent the circuit performance with high efficiency.

The proposed DAWN technique improves surrogate fidelity across backbones by up to +0.055 R^2. However, performance drops to R2 ≈0.281 on unseen topologies, suggesting structural extrapolation across circuit graphs remains a fundamental open challenge.

Critical strengths
  • Circuit dataset consists of simulations across different tech nodes.
  • The proposed method is technically reasonable. The paper includes comprehensive comparisons with different backbones (GCN, GT, ParaGraph, MLP) together with different sizing points and sampling methods.
  • Detailed appendix showing the dataset distribution, sampling method, and testbench configurations.
Critical weaknesses
  • The negative result regarding the lack of transferability to unseen topologies critically limits the practical use of the proposed method.
  • Only OTAs are used in the dataset construction. It is not clear whether the conclusion could be easily extended to other circuit types.
  • Although the model they propose is sound, detailed investigations on how performance varies at each stage. Also, since many performance metrics are used, more experiments can be conducted to show their different weight contributions after training. Are they equally important, or do some metrics dominate the result?
Questions for authors
  1. Other than the global R^2 and Classification Accuracy, are there any other metrics that could be used to evaluate the result?
  2. While the DAWN added could be better than the original architecture, it seems that the architecture itself is also quite important. (0.56 MLP vs 0.7492 GCN) where the architecture difference is larger than DAWN (0.792 vs 0.7929). Did you test different parameters used in the architectures to better show if DAWN's good performance is from its physical insights or just an architecture trick?
  3. Can you elaborate on how technology OOD evaluation is performed? For example, for the same circuit topology, how are parameters of unseen technology nodes provided to the model for inference?
7 comments
AuthorsOpenReview ↗

Response To Reviewer 1ZSL (Introduction, W1, W2)

We would like to thank Reviewer 1ZSL for their insightful review of our work. We are grateful for your recognition of our multi-technology OTA dataset, comprehensive model evaluations, and extensive appendix documentation.

Below, we address each of your stated weaknesses and questions in detail, outlining the specific updates, scope refinements, and newly added tabular statistics added to our revision:

W1: We agree with your assessment regarding the current practical limitations of zero-shot cross-topology transfer. However, within known topologies, our baselines provide substantial practical value by predicting performance across continuous transistor sizing, bias, and environmental sweeps at R² = 0.7929, while successfully characterizing designs across an unseen technology node in 14nm low-standby power (LSTP) at R² = 0.7639. Rather than obscuring structural failure modes, transparently reporting that surrogate prediction drops to R² ≤ 0.2807 on held-out topologies establishes an honest, rigorous baseline that shows where future research on neural surrogate modeling must focus to solve cross-topology transfer within our benchmark. We note that Reviewer J5n7 also supported this perspective in their evaluation of our critical strengths, observing that characterizing both in-distribution successes and out-of-distribution limitations provides meaningful reproducibility assets and an empirical challenge for the research community.

W2: We agree with the reviewer. In response to your critique, we have overhauled the manuscript to eliminate broad generalizations about universal analog scaling laws and strictly scoped our claims to OTAs only:

  1. Title: Explicitly scoped to "ArrivalFM-OTA: A Large-Scale Benchmark for Physical Representation Learning in Operational Transconductance Amplifiers".
  2. Abstract & Section 1 (Pages 1-3): Reframed scope of contribution to a "physical representation learning benchmark comprising over 2 million industrial-grade Cadence Spectre simulation datapoints across 28 operational transconductance amplifier topologies and 10 FinFET technology models from 7nm to 20nm, evaluated under continuous environmental sweeps", and tightened claims of unseen topology and technology process generalization to within the scope of our benchmark itself.
  3. Section 5 (Pages 8-10), Section 6 (Pages 9-10), Appendix I (Pages 57-60): Reanchored to focus on core benchmark diagnostics across the 28 OTA topologies and 10 technology nodes within our benchmark.

Our benchmark thus provides the foundational infrastructure and standardized evaluation splits needed to rigorously study physical representation learning in OTAs. While our current data and evaluation infrastructure is specialized for OTAs only, we view extending this device-level physical representation paradigm modeling both local internal operating states and global circuit performance to other analog circuit classes as vital future work.

AuthorsOpenReview ↗

Response To Reviewer 1ZSL (W3, Q1 Part I)

W3 & Q1: We thank the reviewer for raising these critical points. In the revised manuscript, we have added comprehensive decompositions in Section 5 and Appendix I, providing complete transparency into metric weightings, domain-specific evaluation metrics, and performance variations across stages.

Global Macro-Average Metric Weightings:

We have clarified headline R² and Classification Accuracy metrics as reported in Table 4 (Section 5, Page 9) as unweighted macro-averages in both Table 4 and Appendix G (Pages 49-51) during model evaluation:

  • For macro-averaged global R², each of the 41 directly predicted regression targets in our benchmark receives equal 1/41 weight. Input and output voltage swing ranges are derived bounds (Vmax−VminV_{\text{max}} - V_{\text{min}}) and excluded from macro-averaged global R² weights.
  • For macro-averaged global Classification Accuracy, each of the 10 sequential simulation filters in our benchmark receives equal 1/10 weight.

We find that specific global metrics do contribute asymmetrically to the global R² score. While augmenting GCN, GT, and ParaGraph backbones with DAWN achieves positive ΔR² across all 41 continuous global specifications (and 40 of 41 on MLP), R² gains concentrate heavily in dynamic step-response and stability metrics. In our revised manuscript, we have added Table 32 (Appendix I, Page 59) establishing a full per-metric breakdown of R² gains across our DAWN ablations. Condensed categorical metric summaries of Table 32 is shown below:

Simulation DomainEvaluated SpecsΔR² GCNΔR² GTΔR² ParaGraphΔR² MLP
Small-Signal AC Response7+0.0251+0.0277+0.0398+0.0101
DC Response, Noise, & Periodic8+0.0287+0.0320+0.0448+0.0147
Large-Signal Transient Rise7+0.0381+0.0387+0.0484+0.0183
Large-Signal Transient Fall7+0.0412+0.0410+0.0498+0.0192
Small-Signal Transient Rise6+0.0683+0.0733+0.0722+0.0133
Small-Signal Transient Fall6+0.0703+0.0757+0.0708+0.0156
41-Metric Macro Average41+0.0437+0.0464+0.0532+0.0153
  • Static Metric Saturation: Static DC, gain, and power metrics already achieve high baseline fidelity, where DAWN yields modest, incremental adjustments:
SpecificationUnitEvaluation Count (NN)ΔR² GCNΔR² GTΔR² ParaGraphΔR² MLP
Open-Loop GaindB161,160+0.0289+0.0321+0.0487+0.0035
Unity-Gain BandwidthHz108,869+0.0163+0.0210+0.0328+0.0174
Total PowerW161,160+0.0106+0.0180+0.0283+0.0171
Max Input VoltageV98,772+0.0403+0.0412+0.0591+0.0120
Max Output SwingV98,772+0.0417+0.0389+0.0566+0.0183
Output NoiseVrms101,788+0.0240+0.0277+0.0500+0.0136
  • Stability & Transient Dynamics: Phase margin improves by +0.0786 to +0.1086, while transient offshoot and undershoot metrics improve by +0.0719 to +0.1592:
SpecificationUnitEvaluation Count (NN)ΔR² GCNΔR² GTΔR² ParaGraphΔR² MLP
Phase Margin°108,869+0.0786+0.0866+0.1086+0.0203
Large-Signal Rise Offshoot%81,301+0.0928+0.0719+0.0760+0.0058
Large-Signal Rise Undershoot%81,301+0.0907+0.0878+0.0577+0.0132
Large-Signal Fall Offshoot%75,324+0.1029+0.0834+0.0847+0.0094
Large-Signal Fall Undershoot%75,324+0.0978+0.0985+0.0635+0.0181
Small-Signal Rise Offshoot%96,758+0.1377+0.1579+0.0979+0.0135
Small-Signal Rise Undershoot%96,758+0.1316+0.1356+0.1354+0.0078
Small-Signal Fall Offshoot%97,716+0.1375+0.1592+0.0985+0.0148
Small-Signal Fall Undershoot%97,716+0.1402+0.1454+0.1295+0.0205

Additionally, dynamic phase margin (+0.0786 to +0.1086) and transient offshoot and undershoot percentages (+0.0719 to +0.1592) account for over 56.35% of DAWN GCN's total macro lift (+0.0246 out of +0.0437, where the sum of deltas across all 41 specs is 1.7920, and these 9 specs contribute 1.0098, or (1.0098 / 41) / 0.0437 = 56.35%). While these results reflect the combined benefit of targeted bottleneck capacity and multi-task guidance rather than an isolated physical mechanism, the concentration of gains in phase margin and settling overshoot demonstrates that this architectural augmentation primarily recovers accuracy on the stiff dynamic stability and transient ringing metrics where baseline models struggle most.

AuthorsOpenReview ↗

Response To Reviewer 1ZSL (W3, Q1 Part II)

Stage-Based Investigations & Alternative Evaluation Metrics:

For our DAWN GCN baseline specifically, we report physical domain median absolute error (MedAE), mean absolute error (MAE), and R² for all regression metrics alongside accuracy, precision, recall, and F1\text{F}_1-scores for all classifiers. Equivalent performance statistics for our other 7 baselines can be found as .csv summaries in the Link 3: ArrivalFM-OTA Benchmark Baselines supplementary material, which also contains physical RMSE, Z-score based MAE, and Z-score loss statistics for each regression metric. Each metric in the tables below is reported as mean or mean ± standard deviation across 10 random baseline initializations. We investigate performance across 4 distinct definitions of circuit stages:

(a) Local Component Representation Stage (Table 5, Section 5, Page 10), (Table 31, Appendix I, Page 58):

Internal DC bias states in DAWN GCN are resolved with R² = 0.76 to 0.98 on core device-level parameters, expressed in the table below:

SpecificationUnitEvaluation Count (NN)Physical MedAEPhysical MAER² Score
VdsV_{\mathrm{ds}}mV1,917,76151.20 ± 0.5371.76 ± 0.450.7617 ± 0.0020
VthV_{\mathrm{th}}mV1,917,7615.32 ± 0.267.09 ± 0.230.9879 ± 0.0005
IdsI_{\mathrm{ds}}µA1,917,76135.41 ± 0.73250.53 ± 6.170.7726 ± 0.0025
gmg_mmS1,917,7610.5229 ± 0.01402.4411 ± 0.11560.7626 ± 0.0019
gdsg_{\mathrm{ds}}mS1,917,7610.1655 ± 0.002724.87 ± 3.130.8216 ± 0.0016
CgsC_{\mathrm{gs}}fF1,917,7610.2501 ± 0.00760.4520 ± 0.01350.9672 ± 0.0013
Vnoise,transV_{\mathrm{noise,trans}}µVrms1,084,6793.12 ± 0.1233.59 ± 35.710.7953 ± 0.0059
Vcm,ctrlV_{\mathrm{cm,ctrl}}mV128,93082.83 ± 1.78108.00 ± 1.110.6481 ± 0.0037

Component-level evaluations reveal a clear distinction between parameter classes. Electrostatic variables such as threshold voltage VthV_{\mathrm{th}} and gate capacitance CgsC_{\mathrm{gs}} are resolved with high fidelity. In contrast, operating currents IdsI_{\mathrm{ds}} and conductances gdsg_{\mathrm{ds}} exhibit tight median errors (35.4 µA and 0.17 mS) within the context of larger mean errors (250.5 µA and 24.9 mS). This directly reflects the difficulty of resolving wide-range currents and closed-loop DC operating points in surrogate modeling.

(b) Global Performance Stage (Section 5, Table 6, Page 10):

For our DAWN GCN baseline, evaluating global regression metrics across the In-Distribution split in physical units provides intuitive resolution bounds for design space exploration (5.07 dB DC open loop gain, 4.18° phase margin, and 0.64 µV input offset voltage MedAE), while physical MAE highlights the sensitivity of metrics with large distributions such as unity-gain bandwidth, power, slew rate, and settle time:

SpecificationUnitEvaluation Count (NN)Physical MedAEPhysical MAER² Score
AolA_{\text{ol}}dB161,1605.0666 ± 0.13038.4834 ± 0.13290.9084 ± 0.0034
fugbwf_{\text{ugbw}}MHz108,8699.3706 ± 1.36553066.59 ± 739.680.9588 ± 0.0034
PM\text{PM}°108,8694.1804 ± 0.10168.0708 ± 0.10630.7875 ± 0.0040
CMRR\text{CMRR}dB106,6375.1837 ± 0.36297.2345 ± 0.34790.9854 ± 0.0010
PSRR\text{PSRR}dB106,6375.2906 ± 0.42787.2631 ± 0.44480.9865 ± 0.0012
Vswing,outV_{\text{swing,out}}V98,7720.0768 ± 0.00190.1140 ± 0.00140.8211 ± 0.0029
VOSV_{\text{OS}}µV98,7850.6406 ± 0.113599.29 ± 101.230.8892 ± 0.0031
THDdBc\text{THD}_{\text{dBc}}dB97,7185.1730 ± 0.12766.8346 ± 0.10520.8763 ± 0.0029
PPµW161,160103.37 ± 1.83879.24 ± 23.900.9562 ± 0.0041
Vnoise,outV_{\text{noise,out}}µVrms101,7885.5173 ± 0.178337.63 ± 1.970.9067 ± 0.0035
tsettle,LRt_{\text{settle,LR}}ns81,3017.6343 ± 0.5975(8.39 ± 0.80) × 10⁶0.9701 ± 0.0036
SRLR\text{SR}_{\text{LR}}V/µs81,29913.1146 ± 1.61415805.54 ± 1871.430.9667 ± 0.0034
OSact,LR\text{OS}_{\text{act,LR}}%81,3010.0315 ± 0.00323.0147 ± 0.03840.4960 ± 0.0298
USact,LR\text{US}_{\text{act,LR}}%81,3010.0060 ± 0.00080.0855 ± 0.00110.2765 ± 0.0151

Comments on the paper

2 comments
AuthorsOpenReview ↗

Review Response Summary (Part 2)

4. Architectural & Code Documentation

  • Audited PyG modules and reconciled parameter counts to match instantiated models as graciously noted by Reviewer J5n7 (ParaGraph 5.30M, GT 18.97M, GCN 2.17M, MLP 2.16M); confirmed DAWN adds an identical +148,224 bottleneck to all backbones. Reconciled parameter counts in arguments and conclusions drawn across the Main Text and in Table 27 (Appendix H.1, Page 52)
  • Reconciled manuscript formulas for metric calculation and simulations with frozen code (spec_functions.py, training/*.py):
    • Equation 40 (Appendix E.3.4, Page 38): Corrected phase unwrapping formula.
    • Equations 31, 32, & 33 (Appendix D.9, Page 34): Updated to 20τ20\tau transient settling horizons.
    • Appendix E.3.2 (Page 37): Clarified linear open-loop gain conversion before impedance calculation.
  • Clarified the 2,059,328 benchmark split gathered from running data collection for each of our 28 OTA topologies strictly following Algorithm 1 (Section 3.1, Page 4) in Section 1 and Section 3.2 (Page 5) with the remaining 590,245 datapoints in our dataset represent 100% physically valid Spectre simulations from early exploratory sizing sweeps, which are released alongside our open-source Contact synthesis framework as an unpartitioned archive for open self-supervised pretraining.
  • Power blowup affects 309 of 2,059,328 samples solely in a singular topology: OTA_FD03. DC open loop gains for these 309 samples range from -134.09 dB to -11.15 dB, drawing from 50.06 mW to 976.21 MW of power. We hypothesize that severe voltage headroom starvation in OTA_FD03's 5-transistor stack under unclamped ideal common-mode feedback (CMFB) caused simulator divergence on the auxiliary supply current variable V0:p. These design points are intentionally retained in the benchmark as negative training samples for feasibility classifiers and component-level metric regressors. We have included this explanation of these anomalies in the description of Table 23 (Appendix F, Page 48).
  • Clarified our rationale for introducing different classifier heads in Table 24 (Appendix G, Page 50) to provide insight into our reasoning for introducing classifier heads as a universal diagnostic for defined NaN failures in Appendix D.2 (Pages 31-34) and Appendix E.2 (Pages 36-37). Downstream analyses such as large-signal dynamic swing (swing_stab), DC offset (offset_stab), transient settling envelopes (transient_sr, transient_sf, transient_lr, transient_lf), and periodic steady-state analysis (periodic_stab) are trained to classify an OTA design as invalid (0) if a candidate fails an upstream AC/DC gatekeeper as described in Appendix E.2 or fails to return a valid result given target performance metric definitions in Appendix E.3 (Pages 37-44). Specifically, for periodic steady-state analysis, we require Cadence Spectre's harmonic shooting algorithm to have a stabilization time of tstab=100/fugbwt_{\text{stab}} = 100 / f_{\text{ugbw}}. When unity-gain bandwidth collapses below 1 kHz, the required stabilization window extends to tstab≥100 mst_{\text{stab}} \ge 100\text{ ms}, which we observe commonly causes divergence errors and excessive CPU timeouts. Thus, each head acts as an independent, domain-specific feasibility filter for determining whether a candidate sizing yields valid, non-divergent specifications within that respective characterization domain, while complete end-to-end simulation convergence is independently isolated by the validity head, conditioned on sim_status as defined in Table 21 (Appendix E.2, Page 37). Notably, sim_status retains nominal validity even when simulation stages are declared as invalid under the fugbwf_{\text{ugbw}} ≤ 1 kHz constraint. We also find that high classification performance is preserved across the entire spectrum of class balances. On transient large-signal falling settling (transient_lf), passing prevalence drops to a minority 46.74%, representing the stiffest physical settling constraint in the benchmark. Rather than trivializing predictions toward the majority class, DAWN GCN achieves 92.20% ± 0.08% accuracy, 0.9057 precision, 0.9300 recall, and an F1\text{F}_1-score of 0.9177 on passing designs, while simultaneously detecting failing candidates with 0.9371 precision, 0.9150 recall, and an F1\text{F}_1-score of 0.9259. Conversely, on phase-crossover detection (threshold_stab) where passing prevalence is 95.24% (only 4.76% failing), the model detects rare non-crossing failure states with 0.9199 precision, 0.8923 recall, and an F1\text{F}_1-score of 0.9057. On global convergence (validity, 63.24% pass / 36.76% fail), the model detects simulator non-convergence with 0.9465 precision, 0.9098 recall, and an F1\text{F}_1-score of 0.9278.
AuthorsOpenReview ↗

Review Response Summary (Part 1)

We sincerely thank Reviewers 1ZSL, J5n7, and mU6C for their constructive and rigorous reviews. Their critiques regarding architectural controls, evaluation metrics, dataset curation, architectural framing, and scope have substantially allowed us to elevate the empirical rigor and technical clarity of our manuscript.

Below is a summary of all revisions, refinements, reconciliations and new material in our updated 60-page manuscript:

1. Scope Refinements

  • Paper Title: Explicitly scoped to "ArrivalFM-OTA: A Large-Scale Benchmark for Physical Representation Learning in Operational Transconductance Amplifiers".
  • Abstract & Section 1 (Pages 1–3): Scoped strictly to operational transconductance amplifiers (OTAs) across 28 topologies and 10 FinFET nodes, removing claims of universal analog design. Zero-shot cross-topology transfer within our benchmark (R2≤0.2807R^2 \le 0.2807) is framed transparently as an open challenge.
  • Prior Work: Added LOADBench (Leibl et al., 2026) to Table 1 (Section 2, Page 3); accurately framed INSIGHT (Poddar et al., 2025) in (Section 1, Page 1) as using autoregressive Transformers to predict compute-intensive transient specs from small-signal metrics.

2. Methodological Clarifications

  • Eliminated imprecise phrases describing DAWN ("physics-informed neural networks", "analog waypoints", "physics-conditioned cascade") across the Abstract, Main Text, and Appendix H.2.5 (Page 56)
  • Renamed DAWN as a DCOP-Aware Weighted Network, formalizing it in the Abstract, Section 4.2 (Page 7) and Appendix H.2.5 as an additive multi-task residual regularizer. Added Figure 21 (Appendix H.2.5, Page 56) for a diagrammed visualization of the framework.
  • Clarified in Section 4.2 that zero-initialization preserves the backbone's initial forward mapping and prevents early training destabilization, while auxiliary heads still backpropagate gradients through the identity path.
  • Narrowed our contribution of DAWN to its demonstrated empirical benefit and the zero-initialization guarantee alone. In Section 5, we draw 2 conclusions from experiments detailed in Table 4 (Section 5, Page 8) and Table 32 (Appendix I, Page 59). First, that adding DAWN's +148k parameter bottleneck and auxiliary supervision to our nonrelational MLP baseline yields only +0.0150 R², contrasting with +0.0437 R² across a parameter-matched, graph-based GCN backbone. This demonstrates that while added capacity yields a minor baseline lift, DAWN’s empirical gain is strongly amplified by relational component connectivity. Second, that phase margin and transient ringing metrics account for over 56% of DAWN's macro-averaged lift, primarily recovering accuracy on the stiff feedback stability and step damping targets where baseline models struggle most.

3. Newly Added Empirical Tables & Figures

To further reinforce the empirical rigor of the benchmark, we have added 5 new tables and 2 new scaling figures across Section 5 (Pages 8–10) and a newly instantiated Appendix I (Pages 57–60). These additions provide exhaustive unaggregated evaluations, breaking down internal component-level device operating point metrics (Table 31, Appendix I, Page 58), physical errors in real circuit units (Table 6, Section 5, Page 10), decoupled classification diagnostics across all 10 characterization heads (Table 30, Appendix I, Page 57), cross-quadrant out-of-distribution physical deltas (Table 33, Appendix I, Page 60), and empirical topological complexity scaling correlations with global metric prediction (Figure 3, Section 5, Page 10), (Figure 22, Appendix I, Page 58).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.