- 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?