acceptodds
Under review as a conference paper at ICLR 2027

ThunderGen: A Comprehensive Dataset and Multi-Topology Synthetic Estimator for Fast Analog Circuit Simulation

Abstract

Considering the design of analog integrated circuits (ICs), SPICE-level simulations remain essential for final verification, while their prohibitive computational requirements motivated the research towards synthetic performance estimators to accelerate early-stage design space exploration. However, these surrogates frequently suffer from a significant accuracy gap, failing to capture the non-linearities and high-dimensional complexities of analog circuits. This is mainly due to the lack of high-fidelity datasets covering vast architectural design spaces. Existing works primarily focus on fixed topologies and parameter sweeps, hindering the development of robust machine learning models capable of true topological generalization. This paper presents ThunderGen, a publicly available comprehensive framework and corresponding dataset conceived to bridge the gap between simulation speed and estimation accuracy. ThunderGen introduces a large-scale, multi-topology dataset generated through an automated pipeline that spans diverse circuit families, providing a rigorous ground truth. Alongside this dataset, we propose a performance estimator that serves as a high-fidelity benchmark to replace time consuming SPICE-level simulations and enabling AI-driven topology searches in early design phases. The experimental campaign evaluated ThunderGen against two state-of-the-art methodologies considering a realistic dataset consisting of more than circuits organized in five circuit families and thousands of circuit topologies per each. Experimental results demonstrate that ThunderGen outperforms existing estimators by and respectively, also showing an average relative percentage error compared to SPICE simulation limited to %.

Primary area:
datasets and benchmarks
Keywords:
analog circuits, dataset, simulation, generative AI

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

This paper studies fast machine learning models for analog-circuit performance prediction. ThunderGen uses diffusion models to generate circuit topologies and device parameters, then filters the circuits using Cadence Spectre simulations. The reported dataset contains 44,555 circuits across five analog and RF circuit families. The proposed HybridGNN combines graph features, global descriptors, and separate prediction heads, with an estimator trained for each family. The paper compares the estimators with DiffCkt and FALCON on the generated data and reports an average MPRE of 16.11%, with improvement factors of 4.69 and 6.07. The dataset and pipeline have practical potential, but the interpretation of the accuracy metric and the evidence for unseen-topology generalization need clarification.

Critical strengths

The multi-family dataset and its construction pipeline could be a useful research resource. The work covers five circuit families and seeks to increase connectivity diversity beyond parameter sweeps on fixed topologies. The reported graph representations, netlists, testbenches, and simulation results, together with detailed model and training descriptions, could support more representative benchmarks for performance prediction (Table 3, page 10, and Appendices A.3-A.5.).

Critical weaknesses
  1. The main error calculation does not clearly establish prediction accuracy. Equation (1) is called MAE but averages squared residuals, and it compares specifications with simulated or predicted values. Appendix A.2 describes Table 7 as reporting this error. Some Table 4 entries can be reproduced from Table 7 aggregates: for LNA PowerGain, |14.78 - 15.15| / 15.15 gives approximately 2.44%. Similar aggregate errors relative to a specification do not imply that the prediction matches SPICE for each circuit. If Table 7 instead contains mean performance values, agreement between means is also insufficient. The reported 16.11% and improvement factors therefore remain difficult to interpret. This directly affects the main conclusion. (Equations 1-2 and Tables 4 and 7.)
  2. Generalization to unseen topologies is not sufficiently established. Table 11 gives a 90/5/5 split but does not state whether circuits are grouped by topology. Since Table 3 contains multiple instances per unique topology, different parameter instances of the same connectivity pattern could appear in both training and test sets. Such an evaluation would not establish generalization to unseen structures. The topology definition, overlap between splits, and results on topologies absent from training are needed. This is a validation gap, not evidence that leakage actually occurred. (Tables 3 and 11.)
Questions for authors
  1. What does each Table 7 cell represent: MAE or MSE relative to a specification, or mean performance? Are Table 4 errors calculated per circuit before averaging, or from aggregate values? Please provide errors directly comparing each prediction with its SPICE result, including the relative-error denominator and units. If the issue is a formula or label error, please explain the actual circuit-level calculation so that the headline result can be assessed.
  2. Are the training, validation, and test splits grouped by topology? How is topology identity defined, and how much overlap exists between splits? Please provide results on topologies absent from training, or report the performance on the unseen-topology subset of the current test set.
1 comment

We thank the reviewer for the positive assessment of our dataset and for highlighting areas requiring clearer error metrics and evidence of topological generalization.

Q1) The actual implementation of Equation (1) computed the mean absolute deviation from specification targets, and Equation (2) measured aggregate consistency relative to Spectre deviations. We have corrected Equations (1) and (2) in the revised manuscript. We have also added a comprehensive evaluation reporting direct per-circuit prediction errors against Cadence Spectre ground truth across all 24 individual metrics spanning all five circuit families (Table 14 in Appendix A.8) and comparison with ML baselines evaluated against Spectre (Table 15 in Appendix A.8).

Q2) In our framework, circuits are mapped to attributed component graphs where nodes represent device types and edges represent shared electrical nets (excluding sizing parameters). Two circuits share a topology if and only if their labeled graphs are isomorphic. A topology is unique if it differs by at least one device or connection. We have updated Section 4.1 to better clarify this. Because our test set is drawn from a multi-topology pool (~28k unique topologies across ~44k circuits), 60.6% of the test samples feature topologies completely unseen during training. To address the question directly, we updated Table 4 in the revised manuscript to include a dedicated column evaluating performance strictly on these unseen topologies. On unseen topologies, ThunderGen maintains a robust 19.73% average MPRE, vastly outperforming DiffCkt (73.77%) and Falcon (100.28%). This confirms that ThunderGen learns generalizable circuit physics rather than memorizing familiar schematics.

Reviewer_dP9mOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

ThunderGen combines conditional circuit generation, Spectre validation, and family-specific HybridGNN performance estimators. The reported dataset contains 44,555 circuits across five analog/RF families and 28,012 unique topologies. Using a 90/5/5 split, the paper reports an MPRE averaged across families of 16.11%, compared with 75.56% for DiffCkt and 97.89% for Falcon.

Critical strengths

The pipeline integrates circuit generation, simulation-based validation, dataset construction, and performance estimation.

The dataset spans five circuit families and includes graphs, netlists, testbenches, and simulated metrics, making it a potentially useful research resource.

Appendices A.3–A.5 provide useful architectural, software, and training details.

Critical weaknesses

The reported metric does not establish per-circuit fidelity. As defined in Sec. 4.1 and Appendix A.2, Table 7 aggregates deviations from target specifications, and Table 4 compares these aggregates with the corresponding SPICE aggregates. For example, LNA PowerGain’s 2.44% follows from comparing 14.78 and 15.15 in Table 7. Matching aggregate deviations does not establish agreement on individual circuits. Equation (1) also uses squared errors despite being labeled MAE. The notation and actual computation require clarification; the reported 16.11% and improvement factors do not currently establish estimator accuracy.

Topological generalization is not adequately evaluated. Table 11 specifies a 90/5/5 split with a fixed seed but does not explain topology-level grouping or quantify train–test topology overlap. No separate evaluation on unseen topologies is provided.

Baseline comparisons are insufficiently documented. The manuscript does not adequately describe baseline adaptation, retraining, target processing, or tuning. In particular, it is unclear which DiffCkt component serves as a circuit-to-performance predictor. These omissions prevent a reliable assessment of comparability.

The claimed acceleration lacks measurements. Table 3 reports generation and dataset-validation times, but no estimator latency, throughput, or measured speedup over Spectre is provided. Table 2 describes netlist-derived features, whereas Appendix A.5 additionally mentions operating-point descriptors; the complete computation and runtime cost of the 224 features remain unclear.

Dataset diversity and validation need clearer definitions. The manuscript does not specify topology canonicalization or detailed filtering thresholds. Moreover, Table 3 reports only 479 unique topologies among 6,281 VA circuits, contradicting the claim of thousands of topologies in every family.

Questions for authors

Can you clarify Equations (1)–(2) and report direct per-circuit errors against Spectre for every metric, with appropriate units, normalization, and uncertainty estimates?

What results are obtained under topology-disjoint splits within each circuit family, and how are topology equivalence and train–test overlap determined?

How were DiffCkt and Falcon adapted, retrained, and tuned under comparable data and input conditions?

What is the measured speedup including feature computation, and do any features require simulated operating-point quantities?

What are the family-specific validity criteria, filtering thresholds, and specification-satisfaction tolerances?

1 comment

We thank the reviewer for the constructive feedback and suggestions, which have helped us identify ambiguities in our formulation and strengthened our evaluation.

Q1) The reviewer is entirely correct. The actual implementation of Equation (1) computed the mean absolute deviation from specification targets, and Equation (2) measured aggregate consistency relative to Spectre deviations. We have corrected Equations (1) and (2) in the revised manuscript. We have also added a comprehensive evaluation reporting direct per-circuit prediction errors against Cadence Spectre ground truth across all 24 individual metrics spanning all five circuit families (Table 14 in Appendix A.8) and comparison with ML baselines evaluated against Spectre (Table 15 in Appendix A.8).

Q2) In our framework, circuits are mapped to attributed component graphs where nodes represent device types and edges represent shared electrical nets (excluding sizing parameters). Two circuits share a topology if and only if their labeled graphs are isomorphic. A topology is unique if it differs by at least one device or connection. We have updated Section 4.1 to better clarify this. Because our test set is drawn from a multi-topology pool (~28k unique topologies across ~44k circuits), 60.6% of the test samples feature topologies completely unseen during training. To address the question directly, we updated Table 4 in the revised manuscript to include a dedicated column evaluating performance strictly on these unseen topologies. On unseen topologies, ThunderGen maintains a robust 19.73% average MPRE, vastly outperforming DiffCkt (73.77%) and Falcon (100.28%). This confirms that ThunderGen learns generalizable circuit physics rather than memorizing familiar schematics.

Q3) Both baselines are evaluated under the same ThunderGen test set. FALCON's dataset covers the same five families and the same 20 seed topologies as ours, in the same 45 nm technology, with Cadence-simulated ground truth and the same 16-metric performance vector. We used the pretrained GNN (trained on the entire FALCON dataset) released in the authors' repository.
For DiffCKT, neither code nor pre-trained models is publicly available, so we implemented the GNN performance-estimation component following the description in the paper appendix. DiffCkt describes performance prediction as an output head of the continuous-graph denoising network conditioned on the target performance (in this sense, for baselines comparison in our experiments, we calculate MAE against target specs), so we attached a performance head to our network of the generative pipeline (one model per circuit family). The head runs alongside the continuous diffusion process during circuit generation. Its output on each generated circuit is the DiffCkt prediction for that circuit. We have added a new section in Appendix A.7 to better clarify these aspects and improve reliability of our experiments.

Q4) To quantify the speedup over conventional simulation and prior estimators, we added detailed inference benchmarks in the revised paper (Table 5 in Section 4.3). Due to standard commercial EDA license agreements, publishing absolute wall-clock execution times for proprietary simulators like Cadence Spectre is strictly prohibited. However, while SPICE simulations typically require seconds to minutes per quality metrics evaluation, ThunderGen’s HybridGNN achieves a sub-millisecond evaluation time of 0.53 ms per circuit running entirely on a standard 12-core host CPU without GPU acceleration. Compared under identical CPU execution, HybridGNN is 29.9× faster than Falcon (15.77 ms), and 2,492× faster than DiffCkt’s GPU-based predictor (1.32 s). Also training the models is a fast process. All per-family timing details, including feature extraction overhead, are reported in Table 5.

Q5) Circuits are admitted based on clear validity criteria: successful GPDK045 netlist mapping, basic port connectivity checks, and error-free Cadence Spectre execution. Our filtering thresholds (detailed in Table 13 and Appendix A.6) are strictly conservative sanity bounds used to discard non-converged solver artifacts. We enforce no specification-satisfaction tolerances. Retaining both sub-optimal and compliant designs prevents survivor bias and ensures the model learns true physical boundaries. Grounding the dataset in real, non-divergent physics allows ThunderGen to evaluate both estimation speed and quality directly against real converged Spectre simulations, unlike prior works that benchmark against unvalidated targets and mask severe accuracy degradation. Due to compute limits and compact design, the VA family produced fewer unique topologies and weaker unseen generalization. We plan to expand VA diversity with new generation runs and complex seeds in future work and now explicitly acknowledge this limitation in Section 5.

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

This paper presents ThunderGen, a multi-topology analog circuit dataset. It combines an automated generation pipeline and a generic performance estimator to replace simulation. The resulting dataset contains dataset contains 44,555 validated circuits across five families. Besides, the author(s) propose a HybridGNN model to predict the performance with 16.11% MPRE, better than the baseline methods.

Critical strengths
  1. The problem is well-motivated. SPICE simulation is extremely time-consuming. This paper presents an example with dataset to predict the performance for early design exploration, while keeping SPICE simulation for few final verification.
  2. If the dataset could be open-source, it will be useful for the future research in analog circuit design community. (I cannot find the dataset via the anonymous line now)
Critical weaknesses
  1. The technical novelty is limited. The data generation pipeline is very similar with DiffCkt.
  2. The experiments are not convincing enough. For example, the detailed ablation analysis of GNN predictor and its runtime are missing.
Questions for authors
  1. Please emphasize the technical novelty of this work. For my understanding, compared with the previous work, e.g. DiffCkt, the author(s) claim that ThunderGen includes performance validation with real simulation results.
  2. As the author(s) claim HybridGNN is designed to estimate performance. What’s the inference time of such GNN predictor? Is there any benefit compared with the traditional pipeline?
  3. In Table 4, are the training and testing dataset disjoint at the topology level? If so, how unique topologies are identified and provide results specifically on unseen topologies?
2 comments

We thank the reviewer for the constructive feedback and for recognizing the importance of accelerating analog design exploration.

Q1) While the reviewer is correct that SPICE validation is a key distinction, the novelty of ThunderGen extends significantly beyond that. Unlike prior works that rely on parameter sweeps across a handful of fixed templates (e.g., 20 topologies in Falcon, 28 in DiffCkt), ThunderGen provides a unified framework to train fast, multi-topology performance estimators capable of true structural generalization. Notably, ThunderGen evaluates both (i) performance results in terms of the time required to estimate, and (ii) quality results in terms of the quality of the estimates, directly against real converged SPICE (Cadence Spectre). In contrast, prior estimators are evaluated only against unvalidated targets, which masks severe accuracy degradation (as shown by our competitive baselines in Section 4.2). ThunderGen establishes both the methodology and the first standardized, physically validated benchmark for analog performance prediction. For better clarification, we have added an explicit reference in Section 2 pointing to the detailed discussion in Appendix A.9.

Q2) To quantify the speedup over conventional simulation and prior estimators, we added detailed inference benchmarks in the revised paper (Table 5 in Section 4.3). Due to standard commercial EDA license agreements, publishing absolute wall-clock execution times for proprietary simulators like Cadence Spectre is strictly prohibited. However, while SPICE simulations typically require seconds to minutes per quality metrics evaluation, ThunderGen’s HybridGNN achieves a sub-millisecond evaluation time of 0.53 ms per circuit running entirely on a standard 12-core host CPU without GPU acceleration. Compared under identical CPU execution, HybridGNN is 29.9× faster than Falcon (15.77 ms), and 2,492× faster than DiffCkt’s GPU-based predictor (1.32 s). Also training the models is a fast process. All per-family timing details, including feature extraction overhead, are reported in Table 5.

Q3) We thank the reviewer for this insightful question. In our framework, circuits are mapped to attributed component graphs where nodes represent device types and edges represent shared electrical nets (excluding sizing parameters). Two circuits share a topology if and only if their labeled graphs are isomorphic. A topology is unique if it differs by at least one device or connection. We have updated Section 4.1 to better clarify this. To address the question directly, we updated Table 4 in the revised manuscript to include a dedicated column evaluating performance strictly on these unseen topologies. Because our test set is drawn from a multi-topology pool (~28k unique topologies across ~44k circuits), 60.6% of the test samples feature topologies completely unseen during training. On unseen topologies, ThunderGen maintains a robust 19.73% average MPRE, vastly outperforming DiffCkt (73.77%) and Falcon (100.28%). This confirms that ThunderGen learns generalizable circuit physics rather than memorizing familiar schematics.

Regarding dataset accessibility, the anonymous repository is active at https://huggingface.co/datasets/AnonymUser00/thundergen. The complete archive can be accessed directly from the "Files and versions" tab.

I thank the authors for their clarification. I will keep my score unchanged.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of paper revisions and common responses

We sincerely thank the reviewers (YUxC, dP9m, LnpT) for their constructive assessments and valuable recommendations.

To address the common concerns raised across the reviews, we have revised the paper on these main points:

  • To resolve ambiguities regarding predictive accuracy and metric definitions (dP9m, LnpT), we clarified the formulation in Section 4.1 and introduced extensive per-circuit evaluation directly against Cadence Spectre simulations in Appendix A.8, incorporating normalized errors, ranking correlations, and confidence intervals across all physical targets.
  • In response to requests regarding structural generalization (YUxC, dP9m, LnpT), we formalized our graph-based topology criteria and added dedicated evaluations on the unseen-topology subset in Table 4, demonstrating that the model generalizes robustly to novel schematics.
  • Addressing runtime and practical utility (YUxC, dP9m), we added a comprehensive computational analysis in Section 4.3 that details training convergence, feature extraction overhead, and inference latency, confirming substantial speedups over prior ML estimators.
  • We expanded the appendices to document our physical filtering criteria, detail the baseline adaptation protocols, and better articulate ThunderGen's core contribution in benchmarking multi-topology surrogate models (YUxC, dP9m).

In the revised manuscript, all modifications and new additions are highlighted in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.