acceptodds
Withdrawn from ICLR 2027

WaveFormer: A Behavior-Aware Transformer for Analog Performance Prediction Via Circuit Response Pretraining

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Analog-circuit response curves describe how observed voltages and currents vary with stimulus and physical coordinates, providing supervision beyond individual scalar specifications. We introduce WAVEFORMER, a framework that learns circuit representations from DC, complex AC, and transient responses and transfers its shared circuit–testbench backbone to scalar performance prediction. Shared target-aware circuit and function-aware testbench encoders condition three analysis-specific multiscale response heads, jointly trained with masked value and shape losses. The WAVEFORMER BENCHMARK pairs 33,695 parameter configurations across 30 topologies with response curves and supports evaluation of response fidelity, representation transfer, and model-guided sizing. On held-out configurations of known topologies, WAVEFORMER reduces response normalized root mean squared error (NRMSE) by 25.9–46.5% relative to Deep-ONet. With the same performance-predictor architecture and fine-tuning budget,response pretraining reduces Primary NRMSE by 4.7%, with improvements in DC and AC metrics but a slight degradation in transient metrics. The resulting performance predictor also guides SPICE-verified sizing, connecting response-supervised learning to circuit design.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Physical artificial intelligence (Physical AI), Analog circuits, Circuit representation learning, Response pretraining, Neural operators, Graph Transformers, Circuit simulation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_AiLhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The WaveFormer model acquires a circuit representation via predictions of DC, AC, and transient responses, based on the circuit graph, excitation inputs, and target observations. WaveFormer, in this way, transfers the learned shared backbone for performance prediction in a scalar space. WaveFormer introduces a benchmark comprising 33695 parameter configurations across 30 topologies and shows an improvement in NRMSE from 25.9% to 46.5% compared with DeepONet for unseen configurations of known topologies. Pre-training reduces NRMSE and improves the accuracy of transient predictions.

Critical strengths

S1. WaveFormer addresses a highly relevant problem.
Repeated circuit simulation is a bottleneck in analog design, and accurate performance surrogates could reduce the cost of evaluating candidate designs.

S2. Results show substantial advances.
WaveFormer outperforms all 12 examined response-related metrics relative to the baseline models, achieving lower NRMSE values than DeepONet. Moreover, the dataset consists of 3 analysis types and 30 topologies.

Critical weaknesses

W1. The experimental discussion is insufficiently in-depth.
The experiments do not sufficiently establish which proposed components are necessary across analyses. One key advantage of the pre-training procedure is the reduction in NRMSE. However, it involves additional training compute. The simpler MLP decoder works better than the proposed multiscale decoder for all AC metrics (Table 9). Moreover, there are some baselines with large, unexplained normalized errors (see Table 2). The sizing procedure considers several components (pre-training, ensembling, calibration, search procedures), but their effects are not considered independently. In addition, the comparison with AutoSizer is mixed (Table 6) and not discussed.

W2. The presentation is chaotic.
While individual components are well-defined, the whole pipeline needs to be understood by linking the relationships between observation targets, source representation, response head, scalar metric query, and sizing task in Sections 3-5 and the appendices. Figures 1-3 provide extensive technical information, which is not paired with an equally in-depth description. Differences such as waveform pre-training vs. scalar prediction require substantial reconstruction across the main text and appendix.

Questions for authors

Q1. Clarification. Evidence supporting the main claim.
What is the evidence that makes the value of response supervision distinct from the extra benefit of the 5000 pre-training iterations in Section C.7? Considering Figure 4d, how can one argue for an advantage over scalar-only training given comparable total compute, or is the claim only applicable to the reported specific initializations?

Q2. Clarification. Need for the proposed decoder.
Using an MLP instead of the multiscale decoder in Table 9 improves AC symmetric L2 from 0.203 to 0.182, Pearson correlation from 0.968 to 0.987, NMAE from 0.065 to 0.060, and NRMSE from 0.137 to 0.131. What evidence supports the proposed AC decoder, given the alternative's superiority across all metrics?

Q3. Clarification. Fairness of baseline comparison.
Why do the NRMSE values of AC vary between ~225 and ~1566 (Table 2) for adapted baselines, whereas for DeepONet the value is only 0.256? How is it confirmed that the comparison is valid with respect to model performance?

Q4. Clarification. Design utility.
Which result indicates that response pretraining leads to better sizing results when the ensemble, calibration, optimizer, and simulation budget are kept constant? In the absence of the latter control, and considering the mixed comparison to AutoSizer (Table 6), what is the design utility contribution of the representation learning approach?

Q5. Clarification. Range of conditional learning.
In Section A.1, there is only one target per topology-analysis pair, and all test topologies are present during training. Which variations of excitations and observation targets are being left out? Which outcomes confirm that the model learned to condition on these inputs independently of the parameters?

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

The paper proposes WaveFormer, a model that predicts analog-circuit response curves (DC sweeps, complex AC, transient) from (i) an attributed device–net–matching graph encoded with relation-biased graph attention, (ii) a set-invariant encoder of sampled source/stimulus functions, and (iii) per-analysis DeepONet-style heads with a base trunk plus a Fourier-feature residual trunk. Training uses a masked Smooth-L1 value loss plus slope/curvature/feature "shape" losses.

Critical strengths

strengths:

  • Well-motivated problem: using full DC/AC/transient response curves as supervision, rather than scalar specs only, is a sensible direction, and the pipeline from pretraining to metric prediction to SPICE-verified sizing is complete.
  • Reasonable neural architectural choices for the targeted problems
Critical weaknesses

Weaknesses

  1. The central claim about response pretraining helps is not statistically supported, and most response-prediction baselines are non-functional.

  2. The speedup claim is way overstated, and runtime is the wrong headline for this problem until accuracy is at a level where the surrogate can actually replace SPICE.

(a) The measurement is not a like-for-like comparison. "40–251× faster than simulators" (§5.2, Table 4; 8.682 ms vs. HSPICE 347.6 ms / ngspice 2179.3 ms) compares a batch-1 forward pass on an A100 with weights and inputs already resident, reported as the minimum over 100 timed calls after warm-up, against full CPU simulator executions. Graph construction, feature extraction, normalization and model loading are explicitly excluded from the model side. The ~24k SPICE simulations needed to build the training set, plus pretraining and fine-tuning time, are not amortized anywhere.

(b) SPICE is the ground truth, so a faster wrong answer is not a speedup. Every runtime ratio in Table 4 is computed at unequal accuracy. The surrogate's own numbers show it is not yet a substitute for simulation: DC response NRMSE 0.192 and transient Pearson ρ 0.786 (Table 2); checking on Table 5b, it basically shows that the worst few percent of cells the prediction is off by ~35% of the metric's training range.

(c) From a circuit designer’s point of view, this paper does not really demonstrate the developed model can deliver real impact or somewhat against its motivation claim. Table 4’s simulations take 0.35 - 2.2 s each. And even if they took minutes, a designer would simply run SPICE, with seconds to minutes of simulation are nothing in an actual production flow, and no one would trade exactness for a 250× speedup on a run that already finishes before the results window opens. The case that would make the paper’s motivation real is the opposite one: for example, a large transient simulation, post-layout/parasitic-extracted netlists, PVT corner sweeps and Monte Carlo runs, where HSPICE needs hours per candidate. A surrogate matters only if it can be shown to maintain relatively high prediction precision on exactly those expensive cases, since that is where replacing or pruning simulations changes the design schedule. The paper contains no such case.

Questions for authors

I don't have question to ask.

Reviewer_Vo5WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces WAVEFORMER, a circuit representation learning framework that uses simulated circuit responses as pretraining supervision. Instead of predicting only scalar metrics such as gain, bandwidth, and power, it learns to predict complete DC, AC, and transient response curves based on circuit topology, device parameters, and testbench conditions. The authors also introduce a benchmark with ~33.7k parameter configurations across 30 circuit topologies. Their results show that WAVEFORMER reduces response prediction error significantly compared with DeepONet and improves downstream performance prediction by 4.7%. They also demonstrate its use in SPICE verified multi-objective sizing.

Critical strengths
  • The paper makes a useful contribution by learning from full DC, AC, and transient response curves rather than relying only on scalar circuit specifications.
  • The paper is easy to understand and well-written.
  • The benchmark introduced can serve as a common setting for evaluating response prediction, representation transfer, and sizing.
  • The results are promising, with WAVEFORMER reducing NRMSE significantly over DeepONet. Also, the sizing experiment provides a useful demonstration of design utility.
Critical weaknesses
  • The evaluation uses different parameter configurations of known circuit topologies for training and testing. However, It is unclear if the method can generalize well to entirely unseen circuit topologies.
  • The improvement reported from response pretraining is modest, and the paper doesn't offer sufficient statistical significance testing to establish the reliability of this improvement.
  • The paper doesn't clarify whether the circuit sizing evaluation applies to other circuit types and more complex design tasks as the current setup is limited.
  • The paper must consider the computational cost of generating the simulated response dataset and pretraining the model when discussing the practical benefits over conventional approaches.
  • The results from transient response prediction show a slight dip after pretraining, suggesting that the benefits of response-based pretraining may not be consistent across all circuit analyses.
Questions for authors
  1. How does WAVEFORMER perform on entirely unseen circuit topologies? How do we make them generalizable?
  2. Why does response pretraining improve downstream DC and AC prediction but slightly degrade transient prediction? Are these differences consistent across multiple training seeds?
  3. How much simulated data and computational effort are required to generate the response dataset and pretrain WAVEFORMER?

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

Related papers

Discussion (0)

Sign in to comment.