- 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 layout metrics"? Line 211: what does "Pooling after the two structural blocks" mean? Line 244: what are and ? Line 248: what is the Fuse function? Line 254: what is ? 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
-
Why is the Matern kernel chosen for the GP? How well do other kernels perform?
-
What is the optimization method used to solve equation 14 (line 323)?
-
Do the baselines in the main experiment utilize historical design data or do they start each optimization from scratch?
-
Section 5.2 Neural Response Prediction: did the training data contain 14nm designs?
-
Line 389: I am not sure why AUC-HV is a metric to study. Is it standard?
5 comments
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 , the PDK representation , and the graph readout after the fourth GNN block. The output is the context representation . We have defined this operation immediately after Eq. (5); Appendix D provides the input and output dimensions.
W6. Definition of
In Equation (6), denotes the learned parameters of the MLP that predicts the layout metrics. The projection matrices and are listed separately. The tool configuration is denoted by . 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, : recovers masked, known static graph attributes from graph context, encouraging informative structural representations.
- Topology consistency, : penalizes differences between the topology representations of corresponding circuit instances across PDKs. It acts on before the late FiLM blocks, while the final context 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.
| Objective | ||||
|---|---|---|---|---|
| Full | 1.000 | 1.000 | 1.000 | 1.000 |
| w/o | 1.857 | 1.124 | 1.830 | 1.408 |
| w/o | 1.238 | 1.177 | 1.566 | 1.652 |
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 and 17.7–65.2% without , 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.
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.
| Kernel | log1p MAE@8 ↓ | Spearman@8 ↑ | log1p MAE@16 ↓ | Spearman@16 ↑ |
|---|---|---|---|---|
| Matérn-5/2 | 0.6269 | 0.5911 | 0.4625 | 0.6131 |
| Matérn-3/2 | 0.7312 | 0.5501 | 0.5037 | 0.5776 |
| RBF | 0.7044 | 0.5720 | 0.4989 | 0.5972 |
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.
| Kernel | Normalized AUC-HV ↑ | Final normalized HV ↑ |
|---|---|---|
| Matérn-5/2 | 0.9374 | 1.0000 |
| Matérn-3/2 | 0.9369 | 0.9977 |
| RBF | 0.9358 | 1.0000 |
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 (). 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.
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) with representations of the addressed groups to obtain ex. The context representation zctx encodes the circuit structure and PDK information.
In Eq. (6), learned projections map zctx and ex to a common dimension. The prediction MLP receives the concatenation of zctx, ex, and the elementwise product (Wctxzctx)⊙(Wparex). 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 and βℓ
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 aℓ and bℓ:
\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 $\beta_\ell (z_{\mathrm{PDK}})=0$.