acceptodds
Under review as a conference paper at ICLR 2027

Decoupling Topology and Technology for Physical-aware Analog Tool Parameter Tuning

Abstract

Analog layout generators automate physical implementation, but configuring them for different circuits and process design kits (PDKs) still requires expert knowledge and costly evaluations. Reusing historical tuning experience is difficult because parameter meanings and applicability vary across tasks, while changes in circuit structure and technology affect both layout responses and prediction reliability. We present ADAPT, which decouples circuit topology, PDK information, and parameter semantics into separate representations and fuses them to learn a global neural response prior from historical evaluations: technology information conditions the circuit graph network through feature-wise linear modulation (FiLM), and parameter choices then interact with the joint circuit–PDK context to predict layout metrics. A context-calibrated residual Gaussian process models deviations from this prior and quantifies predictive uncertainty. Target observations update the residual posterior while the neural prior remains fixed, and multi-objective Bayesian optimization selects subsequent tool parameter configurations. We construct a dataset covering 40 transistor-level circuits and three technology setups, comprising 120 circuit–PDK tasks with recorded parameter requests, execution outcomes, and available layout metrics. Experiments show that ADAPT achieves the best hypervolume (HV) any baseline attains in 40 evaluations within an average of 12.5 evaluations, a more than 3 reduction in layout tool executions. Our dataset and code are available online https://anonymous.4open.science/r/ICLR-ADAPT-1CA0/README.md.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Decoupled Representation Learning, Conditional Gaussian Process, Design Space Exploration, Electronic Design Automation, Analog Layout

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_QZ7sOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the analog layout tool parameter tuning problem. It proposes ADAPT, an iterative multi-objective Bayesian optimization pipeline that utilizes a learned Gaussian process prior from historical design data. ADAPT separates circuit topology and its technology from the parameter tuning task. The learned GP prior shows generalization to unseen design tasks. ADAPT achieves higher hypervolume faster than baseline methods.

Critical strengths

This paper presents a careful design for the GP prior model. It covers featurization for circuit design and technology node. The separation between the design feature and the to-be-tuned parameters is a clean choice. The focus on analogy circuit is interesting and provides insights on necessary decisions for this domain.

Across the five test 14nm design task, ADAPT achieves higher hypervolumes overall and leads other baselines on convergence speed. This demonstrates its effectiveness.

Critical weaknesses

The methodology section is poorly written. I understand this paper deals in a specialized domain, however, its current presentation leaves much room for improvement. Some suggestions: Section 4.1: this section should be expanded. For example, explain how "Parameter semantics interact with this context to predict the MM layout metrics"? Line 211: what does "Pooling after the two structural blocks" mean? Line 244: what are γl\gamma_{l} and βl\beta_{l}? Line 248: what is the Fuse function? Line 254: what is ω\omega? Line 265: the two auxiliary terms are relegated to the appendix. Generally if they matter for the main experiment, at least some high-level descriptions are necessary.

The above is a non-exhaustive list for the presentation issue, and I believe they are critical for the quality improvement.

Questions for authors
  1. Why is the Matern kernel chosen for the GP? How well do other kernels perform?

  2. What is the optimization method used to solve equation 14 (line 323)?

  3. Do the baselines in the main experiment utilize historical design data or do they start each optimization from scratch?

  4. Section 5.2 Neural Response Prediction: did the training data contain 14nm designs?

  5. Line 389: I am not sure why AUC-HV is a metric to study. Is it standard?

5 comments
AuthorsOpenReview ↗

Response to the weakness 1-4

We thank the reviewer for the detailed comments and questions. We address each point below.

W1. Writing and clarity

Thank you for identifying these issues. We have revised Sections 4.1–4.3 to clarify the parameter–context interaction, the graph readout after the structural blocks, the FiLM scale and shift, the fusion operation, the prediction-head parameters, and the roles of the auxiliary losses.

W2. How do parameter semantics interact with the context to predict layout metrics?

Here, parameter semantics describe each tool setting's family, selected value, addressed object group, and applicability. The parameter encoder combines the typed parameter view χ(x;c)\chi(\mathbf x;\mathbf c) with representations of the addressed groups to obtain exe_x. The context representation zctxz_{\mathrm{ctx}} encodes the circuit structure and PDK information.

In Eq. (6), learned projections map zctxz_{\mathrm{ctx}} and exe_x to a common dimension. The prediction MLP receives the concatenation of zctxz_{\mathrm{ctx}}, exe_x, and the elementwise product (Wctxzctx)⊙(Wparex)(W_{\mathrm{ctx}}z_{\mathrm{ctx}})\odot(W_{\mathrm{par}}e_x). This interaction allows the predicted effect of a tool setting to depend on the circuit and PDK. The MLP predicts the transformed layout metrics defined in Eq. (7). We have clarified this interaction in Sections 4.1 and 4.3.

W3. Pooling after the two structural blocks

After the second graph block, we take the mean and maximum of the device-node features, and do the same for the net-node features. We concatenate these four 64-dimensional vectors with the log-counts of devices, nets, and connections from device terminals to nets. An affine layer maps the resulting 259 features to the 32-dimensional topology vector. The node features are retained and passed to the next two graph blocks. Pooling therefore produces a summary of the graph without replacing its node features.

W4. Definitions of γℓ\gamma_\ell and βℓ\beta_\ell

They scale and shift the features in graph blocks 3 and 4. For each block, affine transformations of the PDK representation produce two 64-dimensional vectors aℓa_\ell and bℓb_\ell:

\beta_\ell (z_{\mathrm{PDK}})=0.1\tanh(b_\ell).$$ The same scale and shift are applied to every node in that block. The layers producing $a_\ell$ and $b_\ell$ are initialized so that $\gamma_\ell (z_{\mathrm{PDK}})=1$ and $\beta_\ell (z_{\mathrm{PDK}})=0$.
AuthorsOpenReview ↗

Response to the weakness 5-7

W5. Definition of the Fuse function

In Eq. (5), the Fuse function applies a learned affine map to the concatenation of three inputs: the topology representation ztopoz_{\mathrm{topo}}, the PDK representation zPDKz_{\mathrm{PDK}}, and the graph readout Pool(H(4))\mathrm{Pool}(H^{(4)}) after the fourth GNN block. The output is the context representation zctxz_{\mathrm{ctx}}. We have defined this operation immediately after Eq. (5); Appendix D provides the input and output dimensions.

W6. Definition of ω\omega

In Equation (6), ω\omega denotes the learned parameters of the MLP that predicts the layout metrics. The projection matrices WctxW_{\mathrm{ctx}} and WparW_{\mathrm{par}} are listed separately. The tool configuration is denoted by x\mathbf x. We have added this definition immediately after Eq. (6).

W7. Roles of the auxiliary losses

We agree that the main text should explain the roles of both auxiliary terms.

  • Masked-feature reconstruction, Lmask\mathcal L_{\mathrm{mask}}: recovers masked, known static graph attributes from graph context, encouraging informative structural representations.
  • Topology consistency, Ltopo\mathcal L_{\mathrm{topo}}: penalizes differences between the topology representations of corresponding circuit instances across PDKs. It acts on ztopoz_{\mathrm{topo}} before the late FiLM blocks, while the final context zctxz_{\mathrm{ctx}} remains technology-conditioned.

Both terms complement the supervised layout-metric prediction objective in Eq. (8). The matched ablation below evaluates their contributions on five held-out circuits.

ObjectiveAevalA_{\mathrm{eval}}Leq,signalL_{\mathrm{eq,signal}}Leq,powerL_{\mathrm{eq,power}}NcutN_{\mathrm{cut}}
Full1.0001.0001.0001.000
w/o Lmask\mathcal L_{\mathrm{mask}}1.8571.1241.8301.408
w/o Ltopo\mathcal L_{\mathrm{topo}}1.2381.1771.5661.652

Each entry reports the ratio of the three-seed mean MAPE to the corresponding Full value. Lower is better.

Removing either term increases mean MAPE across all four metrics: by 12.4–85.7% without Lmask\mathcal L_{\mathrm{mask}} and 17.7–65.2% without Ltopo\mathcal L_{\mathrm{topo}}, relative to Full. These results support the contribution of both terms to the predictive accuracy of the neural response prior. We have added the high-level descriptions immediately after Eq. (8) in Sec 4.3 and report the ablation in Appendix E.

These gains support shaping the shared prior with reusable structural information and cross-PDK topology consistency, while the final response prediction remains conditioned on the target technology.

AuthorsOpenReview ↗

Response to question 1-3

Q1. Choice of the Matern kernel

The GP models errors left by the neural predictor. We use Matérn-5/2 because these errors need not satisfy the strong smoothness assumption of an RBF kernel. It retains local smoothness while allowing rougher variation [1].

Comparison with other kernels. We compared Matérn-5/2, Matérn-3/2, and RBF using the same neural model, training data, input features, and GP fitting budget. Each kernel was fitted separately. The results below use the same held-out configurations after 8 and 16 attempted target evaluations, with scores averaged equally over five circuits and four metrics.

Kernellog1p MAE@8 ↓Spearman@8 ↑log1p MAE@16 ↓Spearman@16 ↑
Matérn-5/20.62690.59110.46250.6131
Matérn-3/20.73120.55010.50370.5776
RBF0.70440.57200.49890.5972

The low-budget error and ranking measurements assess the residual GP’s role in turning a transferred response prior into predictions that can guide target-specific candidate selection.

We also compared search performance by sequentially revealing existing results from a fixed pool of 64 configurations per circuit, using eight common initial evaluations and 40 attempted evaluations in total. All budgets include failed attempts. The following search scores are averaged over the same five circuits.

KernelNormalized AUC-HV ↑Final normalized HV ↑
Matérn-5/20.93741.0000
Matérn-3/20.93690.9977
RBF0.93581.0000

Matérn-5/2 gives the lowest prediction error and highest rank correlation at both reported budgets. Search performance is similar across the three kernels. These results support our choice of Matérn-5/2 as a practical default. We have added the rationale in Section 4.4 and the comparison in Appendix F.

Q2. The optimization method of equation 14?

We approximately solve Eq. (14) by maximizing qLogNEHVI over a finite candidate set using BoTorch’s optimize_acqf_discrete. At each iteration, we exclude previously queried configurations and screen the remaining candidates using the GP posterior. We then select the configuration with the highest estimated acquisition value from the shortlist (q=1q=1). Section 4.5 describes the candidate construction and the use of BoTorch’s discrete optimizer.

Q3. Historical information used by the baselines

The baselines do not all start from scratch. PPATuner explicitly transfers historical responses between source and target tasks; its original paper formulates this transfer GP and demonstrates transfer between related designs [2]. In our main experiment, PPATuner uses historical parameter–response pairs from the target circuit’s 65 nm instance to optimize its 14 nm instance. This source–target pairing preserves logical circuit identity across technology nodes.

ADAPT’s shared response prior uses explicit circuit and PDK representations to learn across historical circuit–PDK tasks, including 14 nm responses from other circuits.

The other evaluated baselines fit their response models using target-task evaluations. CausalTuner and Reshaping BO additionally use offline pretraining on parameter inputs without historical response labels.

Reviewer_mkQ4OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents ADAPT, a framework that accelerates analog-layout generator tuning across different circuits and PDKs by learning transferable representations of circuit topology, technology, and tool parameters. It combines a neural global prior with an uncertainty-aware residual Gaussian process and Bayesian optimization, reaching the best baseline layout quality in only 12.5 evaluations on average with a 3x reduction in costly tool runs.

Critical strengths
  1. The method explicitly represents circuit connectivity, technology resources, and parameter semantics, including conditional applicability. Combining these representations with a neural response prior and a residual GP provides a well-motivated mechanism for reusing historical experience while adapting to a target task. This addresses a consequential practical limitation of optimizing each circuit–PDK pair independently.

  2. Across five benchmarks and three optimizer seeds, ADAPT reportedly reaches the strongest baseline’s final hypervolume in approximately 12.5 evaluations on average, versus a 40-evaluation baseline budget. Counting initialization and unsuccessful evaluations makes this a meaningful measure of tuning efficiency. The accompanying dataset of 120 circuit–PDK tasks is also a potentially useful research contribution.

Critical weaknesses
  1. The reported Pareto fronts require connectivity-qualified layouts but do not require zero DRC violations. Although some examples are reported as DRC-clean under ALIGN’s checker, this does not establish the validity of all layouts contributing to the optimization results. Furthermore, no explicit layout-versus-schematic verification is reported; checks for opens and shorts alone do not establish device and parameter correspondence. Consequently, the reported improvements are not demonstrated over consistently DRC-clean, LVS-correct implementations.

  2. Missing pre vs post layout metrics discussion. The evaluation considers area, equivalent metal lengths, and via-cut count, without reporting parasitic-extraction-based simulations or discussing electrical performance before and after layout. These geometric objectives do not establish preservation of circuit specifications, particularly when tuning changes device implementation, diffusion sharing, and routing. Representative optimized layouts should therefore be evaluated against their schematic counterparts using circuit-specific metrics, such as comparator delay and offset or buffer settling and stability, to establish whether the geometric improvements preserve electrical performance.

Questions for authors
  • LVS and DRC checks are weak. Indeed not all the produced layouts are valid. How can the flow manage such scenarios?
  • quality results should report both area and functional quality metrics to show the fit with circuit specifications. is the tool capable of considering both area and functional quality metrics?
4 comments
AuthorsOpenReview ↗

Response to Weakness 1 & Question 1

We thank the reviewer for recognizing ADAPT’s transfer mechanism and evaluation efficiency. We have supplemented the evaluation to address the physical validity of the results and their electrical implications. Requiring zero reported DRC violations retains 99.87% of ADAPT’s mean HV; device-level LVS now covers implementations of all five benchmark circuits; and circuit-specific pre-layout/post-layout measurements and baseline comparisons are provided. We address the two weaknesses and the associated questions below. The corresponding results are now reported in Appendix G (Tables 17–18) and Appendix J (Tables 21–23).

Supporting evidence. Native DRC/LVS reports, pre-layout/post-layout simulation results, and waveform-based functional checks for the five supplementary circuit implementations are available in our anonymous evidence package, with per-circuit records listed in the file index.

W1 & Q1. DRC qualification, device-level LVS, and invalid candidates

ADAPT’s achieved geometric quality is largely retained under the stricter DRC requirement. Following Appendix A.1, C denotes completed full-flow observations with all four finite post-route metrics and no opens, shorts, or unresolved connectivity/checker diagnostics. L additionally requires a completed check under the frozen ALIGN checker with zero reported DRC violations. The original optimization fronts use C.

Of ADAPT’s 116 archived frontier entries, 112 already satisfy L. Reconstructing the fronts from all L-qualified observations, with unchanged objective normalization and reference point, retains 99.87% of the original mean HV: 0.0526592 versus 0.0527291. Thus, excluding observations that fail the stricter DRC qualification has a negligible effect on ADAPT’s achieved geometric front quality.

The L-qualified evaluation counts for ADAPT and all baselines are:

MethodL-qualified / evaluated requests
ADAPT460 / 600
DAC’20433 / 600
PPATuner430 / 600
Attention-EDA429 / 600
CausalTuner213 / 600
Reshaping BO456 / 600

The practical implication is that the geometric quality obtained through ADAPT’s experience reuse is largely preserved when the declared DRC requirement is enforced.

Invalid candidates are excluded from the qualified results, while their evaluation costs remain counted. In the supplementary search with L qualification, only revealed L-qualified observations update the residual GP posterior and enter the acquisition baseline and Pareto front. Candidates with failed or incomplete required checks are excluded from these updates; missing checks are not interpreted as zero violations. All requests, including initialization and unsuccessful evaluations, remain charged to the budget. Consequently, favorable geometric metrics cannot improve the accepted front when a candidate fails the required qualification.

Device-level verification now covers all five benchmark circuits. We performed Calibre DRC and device-level LVS on implementations of CLK-CMP, CT-CMP, DC-Switch, PT-CMP, and REF-BUF. All five report zero violations under the same 49-rule research deck and an LVS result of CORRECT against their generator-expanded reference netlists.

CircuitCalibre DRC violationsDevice-level LVSMatched MOS groups / nets / named ports
CLK-CMP0CORRECT162 / 153 / 7
CT-CMP0CORRECT62 / 53 / 7
DC-Switch0CORRECT56 / 53 / 13
PT-CMP0CORRECT141 / 128 / 11
REF-BUF0CORRECT108 / 93 / 9

The declared DRC rules cover poly/fin width, metal width and spacing, via-cut dimensions and spacing, and via enclosure. The reference netlists explicitly represent the implemented device geometry and generator-added edge-device groups. LVS checks device types, connectivity, named ports, and geometric L/W with zero tolerance. The table reports objects after parallel-MOS reduction; no ports, nets, or device groups remain unmatched.

These results directly address device and parameter correspondence for the verified implementations, beyond the original connectivity checks. Native Calibre DRC summaries and device-level LVS reports are provided in Evidence/DRC/<circuit>/Calibre/drc.summary and Evidence/LVS/<circuit>/device_lvs.rep, respectively; the corresponding per-circuit records are listed in the evidence file index. The two additional archived CT-CMP and PT-CMP examples also report zero Calibre violations under the same 49-rule set, as recorded in their respective Evidence/DRC/<circuit>/ALIGN/calibre_drc.summary files.

These checks provide a device-level basis for interpreting the supplementary electrical results, connecting the generated layouts to their declared circuit implementations.

AuthorsOpenReview ↗

Response to Weakness 2 & Question 2

W2 & Q2: Pre-layout/post-layout electrical performance and area

The supplementary evaluation now reports circuit-specific electrical measurements for all five circuits. Simulations use Spectre with PTM-MG 14 nm HP predictive models at 0.8 V and 27°C. For each supplementary implementation, the pre-layout control uses the same device decomposition, sizes, models, bias, stimulus, and load as its post-layout counterpart, while omitting layout-derived interconnect RC. The post-layout RC evaluation includes geometry-derived distributed metal/via resistance and ground capacitance, including supply and ground interconnects.

CLK-CMP, CT-CMP, and PT-CMP use 10 fF output loads. CT-CMP uses VBN/VBP = 0.25/0.55 V. DC-Switch uses 10 fF per output and externally driven 0.8/0.4/0 V rails. REF-BUF uses a 20 µA bias, a 100 kΩ load in parallel with 1 pF, and a 20 mV input step. For the supplementary PT-CMP implementation, the same sensing-inverter sizing adjustment is used in both conditions.

CircuitElectrical metricPre-layout controlPost-layout RC
CLK-CMPMean decision time (ps)51.90190.01
CT-CMPLow / high output (V)0.01394 / 0.714770.13686 / 0.64310
CT-CMPRising / falling signed lag (ps)47.98 / −17.69106.84 / 87.16
CT-CMPDC trip differential input (mV)32.8046.515
DC-SwitchMaximum sampled rail error (mV)3.1074.116
PT-CMPRising / falling mean signed lag (ps)184.84 / 10.49418.67 / 77.99
REF-BUFStep gain0.747280.73759
REF-BUFRising / falling settling time (ns)1.955 / 3.0764.465 / 6.106
REF-BUFMaximum DC tracking error (mV)28.96426.658
REF-BUFLocal phase margin (deg)61.7953.96
REF-BUFLoop unity-gain frequency (MHz)757.46374.60

CLK-CMP decision time is measured from the clock 50% crossing until complementary 80%/20% output levels remain satisfied. CT-CMP signed lag is measured from the differential-input zero crossing to the output 0.4 V crossing; the negative pre-layout falling value follows from this crossing convention. Its DC trip measurement gives the nominal differential-input switching threshold. PT-CMP timing uses the corresponding input 50% crossing. REF-BUF settling uses 1% of the actual output-step amplitude, and local loop measurements probe VFB1 with VFB2 closed.

The circuit-dependent parasitic shifts highlight the importance of retaining circuit-specific evaluation within a framework that reuses tuning knowledge across heterogeneous analog designs.

All listed nominal waveform requirements are met in both the pre-layout and post-layout RC evaluations. The criteria and outcomes are:

CircuitEvaluated waveform requirementPre-layoutPost-layout RC
CLK-CMPSix selected clocked decisions with complementary high ≥ 0.64 V and low ≤ 0.16 V6 / 66 / 6
CT-CMPSettled low < 0.16 V and high > 0.64 VPassPass
DC-SwitchTwenty sampled output-rail checks with absolute error < 0.1 V20 / 2020 / 20
PT-CMPEight selected input words with correct adjacent-pair equality outputs and high ≥ 0.64 V / low ≤ 0.16 V8 / 88 / 8
REF-BUFPositive settled output response to a positive input stepPassPass

All 36 recorded checks pass in each condition. The same checks also pass in the archived coupling-sensitivity condition, giving 108/108 checks across the three conditions. In the RC evaluation, REF-BUF retains 98.70% of its matched pre-layout step gain, with a measured local phase margin of 53.96°. The pass/fail results refer to the listed waveform requirements; the accompanying measurements quantify timing, tracking accuracy, settling, and local loop response. Detailed measurements and raw simulator outputs are provided under Evidence/Simulation/, with per-circuit records listed in the evidence file index. The files metrics.csv, pre_post_comparison.csv, and functional_observations.csv provide the electrical metrics, pre-layout/post-layout comparisons, and waveform-check results, respectively.

Explicit circuit-specific acceptance criteria connect the supplementary layout evaluation to the nominal behavior of each design, giving efficient analog-layout tuning a concrete functional context.

AuthorsOpenReview ↗

Response to Weakness 2 & Question 2 (continued)

Area and matched baseline comparisons

Appendix G, Table 17 reports area results together with a separate matched electrical comparison. The following areas come from the main optimization experiment:

Circuit and baselineBaseline area (µm²)ADAPT area (µm²)Reduction
CT-CMP, DAC’20137.10108.6320.77%
REF-BUF, Reshaping BO95.9078.7617.87%

These area reductions connect ADAPT’s reuse of tuning experience to a direct physical-design benefit across two circuit types, complementing the aggregate hypervolume evaluation.

For the supplementary matched electrical comparison, we tested the CT-CMP implementations from ADAPT and DAC’20 and the REF-BUF implementations from ADAPT and Reshaping BO. All four report zero Calibre DRC violations under the declared 49-rule set and device-level LVS CORRECT against their generator-expanded references, including zero-tolerance geometric L/W checks.

These matched implementations are reported separately from the five-design supplementary evaluation above. Their electrical comparisons use the same circuit-specific operating conditions described above.

CT-CMP implementationLow / high output (V)Rising / falling signed lag (ps)
DAC’20, post-layout0.11916 / 0.58967160.18 / 73.49
ADAPT, post-layout0.10677 / 0.60349130.31 / 75.48

ADAPT provides a larger CT-CMP output swing and an 18.65% lower rising signed lag than DAC’20. Both rising and falling measurements are reported.

The CT-CMP comparison illustrates how reusing layout-tuning experience can yield a circuit-relevant timing benefit, even when the search objectives are geometric.

REF-BUF implementationStep gainRising / falling settling time (ns)Local phase margin (deg)Loop unity-gain frequency (MHz)
Reshaping BO, post-layout0.742174.895 / 6.06649.36407.73
ADAPT, post-layout0.616394.925 / 6.02699.09139.00

ADAPT’s rising and falling settling times differ from Reshaping BO by less than 0.7%, and its local phase margin is larger. The corresponding step gain and loop unity-gain frequency are reported alongside these measurements to make the electrical tradeoffs explicit.

The similar settling times support the practical relevance of the tuned implementation, with gain, bandwidth, and local stability defining the accompanying electrical tradeoff.

Can ADAPT consider area and functional quality together?

Yes, at the framework level. Section 3 defines an M-dimensional evaluator response, so simulator-measured delay or settling time can be incorporated alongside area. Circuit-specific electrical limits can also define acceptance conditions together with physical verification. Adding electrical objectives requires fitting the corresponding response models to electrical measurements.

The reported BO experiments optimize evaluated layout area, signal and power equivalent metal lengths, and intermetal via-cut count. The supplementary search applies L qualification, while the added LVS and simulation studies provide direct device-level and electrical assessment. This distinguishes the demonstrated optimization results from the additional verification and makes the measured electrical behavior explicit.

We are especially grateful for your careful review and constructive guidance, which have helped us strengthen ADAPT’s physical and electrical validation, and sincerely hope the additional evidence supports a more favorable assessment of its contribution to reducing costly layout evaluations through a shared neural response prior and target-specific residual correction; we would greatly welcome further discussion of any remaining concerns.

Dear Reviewer mkQ4,

We are the authors of “Decoupling Topology and Technology for Physical-aware Analog Tool Parameter Tuning.” Thank you again for your careful review. Your comments have been particularly helpful in strengthening our physical and electrical validation.

Our responses now show that requiring zero reported DRC violations retains 99.87% of ADAPT’s original mean HV. We also provide device-level LVS for the supplementary implementations, pre-layout/post-layout electrical measurements, and matched baseline comparisons, and clarify how invalid candidates are excluded while their evaluation costs remain counted. We hope this evidence strengthens your confidence in the practical value of ADAPT’s approach to reducing costly layout evaluations.

We would sincerely appreciate your feedback on whether these additions address your main concerns. If they do, we would be grateful if you could consider a more favorable rating. Please let us know if any concerns remain; we would welcome further discussion and the opportunity to clarify them.

Thank you again for your time and constructive guidance.

Best regards,
The Authors

Reviewer_pS2FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed ADAPT, which can transfer experience in analog layout tuning across circuits and PDKs by decoupling topology, technology, and parameter semantics, learning a neural response prior, and adapting a context-calibrated residual GP to target observations.

Critical strengths

The paper addresses a critical problem at analog layout stage: reuse historical tuning experience to reduce expensive analog layout evaluations across PDKs. The experiments also cross 3 different PDKs with 50 different circuits.

Critical weaknesses

Lack of DRC and LVS checks for the layout task. For analog physical design tasks, DRC & LVS checks should first be passed, and then the specs should be optimized. However, the authors only report optimizing area, signal and power metal lengths, and intermetal via-cut count. The missing parts are DRC and LVS checks, how the specs change after layout, and whether these specs still meet the requirements.

Questions for authors

Please clarify the original format and the source of the PDK data, and the methods mapping these raw input to latent representation z. Since there are only three PDKs, what happens if the latent representation z is replaced with a numerical PDK ID? Please provide evidence showing that the PDK encoding is useful.

Please provide DRC and LVS results after optimization, as well as changes in the specs. Please also compare the DRC and LVS results and post-layout specs with those of the baselines.

The authors claim that ADAPT can transfer analog layout tuning experience across circuits and PDKs. However, all optimization experiments are conducted on five 14 nm circuits, while the 14 nm PDK is already represented in the training data. Therefore, it remains unclear whether the method supports transfer to an unseen PDK. Please provide a leave-one-PDK-out experiment, for example, training only on SKY130 and 65 nm and then transferring to 14 nm.

Are 14-nm and 65-nm PDKs open source? If not, how do we reproduce the framework?

7 comments
AuthorsOpenReview ↗

Response to question 1

We thank the reviewer for recognizing the value of reusing analog layout tuning experience across circuits and PDKs, and for highlighting the need for further evidence on physical verification and transfer to unseen PDKs. Below, we address PDK encoding, DRC/LVS and electrical performance, leave-one-PDK-out transfer, and reproducibility.

Q1. PDK data sources, encoding, and comparison with numerical PDK IDs

We use ALIGN's publicly available FinFET14nm_Mock_PDK and Bulk65nm_Mock_PDK, together with the public SKY130 adapter. Metal-layer and via rules are obtained from structured technology descriptions such as layers.json, while device and generator information comes from the corresponding PDK resources. The official ALIGN documentation describes the JSON format used for these technology rules.

As described in Section 4.2, Eq. (3), we organize these inputs into metal-layer, via, and process/generator tokens. These describe layer order and routing rules, via geometry and enclosure, and device families and generator capabilities, respectively. Via endpoints are retained as token attributes. Presence indicators distinguish missing attributes from physical zeros. Type-specific encoders and mean/max pooling produce representations that are concatenated and mapped to the PDK representation. This representation conditions the later GNN blocks through FiLM and contributes to the final context representation.

Following your suggestion, we added a comparison of physical descriptors, scalar PDK IDs, and categorical PDK IDs for response prediction across four random seeds (Appendix I, Table 20):

SeedPhysical descriptors: MAPE (%)Scalar ID: MAPE (%)Categorical ID: MAPE (%)
NO18.39679.22479.5536
NO28.48539.57399.6719
NO38.921810.09989.2780
NO48.79579.56969.4253
Mean8.64999.61709.4822

Based on the mean MAPE across the four seeds, physical descriptors reduce prediction error by 10.06% relative to scalar IDs and 8.78% relative to categorical IDs, with lower error than both controls for every seed. These results support the value of physical PDK information for response prediction.

Physical technology attributes thus serve as meaningful conditioning information for the shared response prior, linking cross-task tuning knowledge to the constraints governing layout generation.

AuthorsOpenReview ↗

Response to question 2

Q2. DRC/LVS results and post-layout electrical performance

These results are now reported in Appendix G (Tables 17–18) and Appendix J (Tables 21–23).

We supplemented the evaluation with a DRC audit across ADAPT and all baselines, device-level verification and pre-layout/post-layout electrical measurements covering all five benchmark designs, and matched CT-CMP and REF-BUF comparisons with the baselines. The five-design supplementary evaluation and the matched baseline comparison are reported separately below because they concern different implementations.

Supporting evidence. Native DRC/LVS reports, pre-layout and post-layout simulation results, and waveform-based functional checks for the five supplementary circuit implementations are available in our anonymous evidence package. The file index lists the corresponding records for each circuit.

DRC results across methods. Following Appendix A.1, C denotes observations with a complete full flow, all four finite post-route metrics, and no opens, shorts, or unresolved connectivity/checker diagnostics. L is the subset of C that additionally completes checking under the frozen ALIGN checker with zero reported DRC violations. The original optimization fronts use C observations.

The numbers of L-qualified evaluations are:

MethodL-qualified / evaluated requests
ADAPT460 / 600
DAC'20433 / 600
PPATuner430 / 600
Attention-EDA429 / 600
CausalTuner213 / 600
Reshaping BO456 / 600

For ADAPT, 112 of 116 archived frontier entries satisfy L. Reconstructing its fronts from all L-qualified observations, with unchanged objective normalization and reference point, retains 99.87% of the original mean HV: 0.0526592 versus 0.0527291. Thus, the stricter DRC qualification has a negligible effect on the geometric quality achieved by ADAPT.

The practical implication is that the geometric quality obtained through ADAPT’s experience reuse is largely preserved when the declared DRC requirement is enforced.

Invalid candidates during search. In the supplementary search with L qualification, only L-qualified observations enter the Pareto front and acquisition baseline and update the residual GP posterior. Candidates with failed or incomplete required checks are excluded from these updates. All requests, including initialization and unsuccessful evaluations, remain charged to the budget; missing checks are not treated as zero violations. Consequently, invalid candidates cannot improve the qualified front merely through favorable geometric metrics, and their evaluation costs remain included.

Device-level LVS across all five designs. We provide supplementary Calibre DRC and device-level LVS results for CLK-CMP, CT-CMP, DC-Switch, PT-CMP, and REF-BUF. All five supplementary implementations report zero violations under the same 49-rule research deck and an LVS result of CORRECT against their independently constructed generator-expanded reference netlists.

CircuitCalibre DRC violationsDevice-level LVSMatched MOS groups / nets / named ports
CLK-CMP0CORRECT162 / 153 / 7
CT-CMP0CORRECT62 / 53 / 7
DC-Switch0CORRECT56 / 53 / 13
PT-CMP0CORRECT141 / 128 / 11
REF-BUF0CORRECT108 / 93 / 9

The declared rules cover poly/fin width, metal width and spacing, via-cut size and spacing, and via enclosure. The reference netlists explicitly represent the generator's implemented device geometry and edge-device groups. LVS checks device types, connectivity, named ports, and geometric L/W with zero L/W tolerance; the table reports matched objects after parallel MOS reduction. Thus, these implementations are DRC-clean under the specified rule set and LVS-correct against the stated references within the declared checking scope. Native Calibre DRC summaries and device-level LVS reports are provided in Evidence/DRC/<circuit>/Calibre/drc.summary and Evidence/LVS/<circuit>/device_lvs.rep, respectively, with execution logs alongside them; the per-circuit records are listed in the evidence file index.

These checks provide a device-level basis for interpreting the supplementary electrical results, connecting the generated layouts to their declared circuit implementations.

AuthorsOpenReview ↗

Response to question 2 (continued)

Pre-layout and post-layout electrical behavior on all five designs. Simulations use Spectre with the public PTM-MG 14 nm HP predictive models, a 0.8 V supply, and a temperature of 27°C. For each supplementary implementation, the pre-layout control uses the same device decomposition, sizes, models, bias, stimulus, and load as the corresponding post-layout simulation, with layout-derived interconnect RC omitted. Post-layout simulations include geometry-derived distributed metal/via resistance and ground capacitance, including supply and ground interconnects.

CLK-CMP, CT-CMP, and PT-CMP use 10 fF output loads; DC-Switch uses a 10 fF load on each of its four outputs with externally driven 0.8/0.4/0 V rails. CT-CMP uses VBN/VBP = 0.25/0.55 V. REF-BUF uses a 20 µA bias, a 100 kΩ load in parallel with 1 pF, and a 20 mV input step. For the supplementary PT-CMP implementation, both the pre-layout and post-layout conditions use the documented two-transistor sensing-inverter sizing adjustment.

CircuitElectrical metricPre-layout controlPost-layout RC
CLK-CMPMean decision time (ps)51.90190.01
CLK-CMPMean VSUP power (mW)0.272000.15940
CT-CMPLow / high output (V)0.01394 / 0.714770.13686 / 0.64310
CT-CMPRising / falling signed lag (ps)47.98 / -17.69106.84 / 87.16
CT-CMPDC trip differential input (mV)32.8046.515
CT-CMPMean VSUP power (mW)9.820071.80135
DC-SwitchMax sampled rail error (mV)3.1074.116
DC-SwitchMax upper overshoot (mV)21.62413.069
DC-SwitchMean VHIGH power (µW)2.648082.74220
PT-CMPRising / falling mean signed lag (ps)184.84 / 10.49418.67 / 77.99
PT-CMPMean VSUP power (mW)0.082320.07274
REF-BUFStep gain0.747280.73759
REF-BUFRising / falling settling time (ns)1.955 / 3.0764.465 / 6.106
REF-BUFMax DC tracking error (mV)28.96426.658
REF-BUFLocal phase margin (deg)61.7953.96
REF-BUFLoop unity-gain frequency (MHz)757.46374.60
REF-BUFMean VSUP power (mW)0.187960.12396

CLK-CMP decision time is measured from the clock 50% crossing until complementary 80%/20% output levels remain satisfied. CT-CMP signed lag is measured from the differential-input zero crossing to the output 0.4 V crossing; its negative pre-layout falling value follows from this crossing convention. PT-CMP signed lag uses the corresponding input 50% crossing as its timing reference. REF-BUF settling uses a tolerance of 1% of the actual output-step amplitude, and local loop measurements probe VFB1 with VFB2 closed. Power entries identify the measured supply source; DC-Switch's VHIGH entry is the power supplied by that rail.

The circuit-dependent parasitic shifts highlight the importance of retaining circuit-specific evaluation within a framework that reuses tuning knowledge across heterogeneous analog designs.

Satisfaction of the declared functional requirements. Under the reported testbench conditions, all five supplementary implementations satisfy the nominal functional requirements listed below.

CircuitRecorded response criterionPre-layoutPost-layout RC
CLK-CMPSix selected clocked decisions with complementary high ≥ 0.64 V and low ≤ 0.16 V6 / 66 / 6
CT-CMPSettled low < 0.16 V and high > 0.64 VPassPass
DC-SwitchTwenty sampled output-rail checks with absolute error < 0.1 V20 / 2020 / 20
PT-CMPEight selected input words with correct adjacent-pair equality outputs and high ≥ 0.64 V / low ≤ 0.16 V8 / 88 / 8
REF-BUFPositive settled output response to the positive input stepPassPass

All 36 recorded response checks pass in each condition. The same checks also pass when the RC network is augmented with the archived nominal-track coupling estimate, giving 108/108 checks across the three conditions. In the RC evaluation, REF-BUF retains 98.70% of its matched pre-layout step gain, with a measured local phase margin of 53.96°. The pass/fail conclusion applies specifically to the waveform criteria listed above; timing, tracking accuracy, settling, power, and local stability are reported quantitatively at the stated operating points. Measurement definitions, numerical results, and raw Spectre outputs are provided under Evidence/Simulation/, as documented in the evidence guide. For each circuit, metrics.csv, pre_post_comparison.csv, and functional_observations.csv provide the measured metrics, pre-layout/post-layout comparisons, and functional-check results, respectively; all records are listed in the file index.

Explicit circuit-specific acceptance criteria connect the supplementary layout evaluation to the nominal behavior of each design, giving efficient analog-layout tuning a concrete functional context.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revisions and Contributions: A Request for Further Consideration and Discussion

Dear Area Chair and Reviewers,

We sincerely thank you for your time and constructive feedback. We have clarified the method and evaluation protocols, added component ablations and PDK-encoding comparisons, and evaluated transfer to an unseen PDK. We have also supplemented the evaluation with DRC/LVS verification and pre-layout/post-layout electrical measurements, including baseline comparisons.

For the ICLR and broader AI community, ADAPT offers a new formulation of a practically important transfer-learning problem: reusing optimization experience when circuit structure, technology constraints, and parameter semantics vary across tasks. It learns a shared neural response prior from separate representations of these factors and uses a residual GP to adapt predictions with limited target observations. Our evaluation examines transfer across circuit designs and PDKs, including transfer to a 14 nm technology setup.

We view this work as a step toward more reliable and generalizable tuning of automated analog placement and routing. Learning to reuse tuning experience could reduce the repeated manual adjustment required from experienced engineers for weeks and support the development of industrial analog design tools that adapt more efficiently to new circuits and processes. This connects physically informed representation learning with a practical need in circuit design and provides a direction for future research and deployment. Our public dataset of 120 circuit–PDK tasks provides a resource for further study.

We also appreciate the strengths highlighted by the reviewers. Reviewer pS2F emphasized the importance of reusing historical tuning experience to reduce expensive analog layout evaluations across PDKs. Reviewer mkQ4 recognized the well-motivated combination of explicit representations, a neural response prior, and a residual GP for experience reuse and target adaptation, as well as the meaningful efficiency evaluation that counts initialization and unsuccessful evaluations, and the dataset’s potential research value. Reviewer QZ7s highlighted the careful representation design and the clean separation between design features and tuning parameters, and noted the higher hypervolume and faster convergence in the reported experiments.

We would be grateful for your consideration of the additional evidence, the clarified technical contribution, and the broader potential of this research direction. Thank you again for your time and guidance. We welcome further questions and discussion.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.