acceptodds
Under review as a conference paper at ICLR 2027

FabBench: A Fabrication-Informed Benchmark for SEM-Domain Process-Conditioned Lithography Models

Abstract

Accurately predicting semiconductor lithography patterns and defect states under continuous process parameters is vital for advanced chip design. However, existing benchmarks either focus purely on physics simulation or isolate defect analysis from SEM domain observations. The scarcity of process-annotated wafer data further prevents joint evaluation of geometric evolution and defect formation under continuous process variations, leaving their coupled behavior largely unexamined. Moreover, conventional image-level metrics are poorly aligned with fabrication-side assessment, often obscuring localized metrological errors and failures in defect prediction. To address these limitations, we introduce **FabBench**, the first comprehensive benchmark for evaluating process-conditioned geometric evolution and defect formation in the SEM domain. Built on 547K physics-grounded, real-SEM-informed triplets with structured in-distribution and out-of-distribution splits, FabBench establishes a hierarchical, metrology-aware evaluation protocol spanning visual fidelity, localized dimensional accuracy, and topological defect consistency. Extensive evaluations show that predictions with high image-level similarity can still exhibit substantial errors under fabrication-aligned metrology and defect metrics, revealing a critical gap between visual fidelity and fabrication-relevant accuracy. FabBench thus provides a unified testbed for developing models that accurately capture both fine-grained geometric responses and topological defect formation across continuous process variations.

Primary area:
datasets and benchmarks
Keywords:
benchmark, dataset, evaluation metrics, conditional lithography model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SvEjOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper builds a benchmark for predicting SEM images of printed lithography patterns from a mask and process conditions (illumination source, defocus, dose). Because real SEM data with dense process labels is hard to obtain, the targets are synthetic: a rigorous lithography simulator produces 3D resist profiles, and a ControlNet trained on real CD-SEM images renders the top/bottom contours into SEM-like images. This gives 547K mask/process/image triplets, split into ID (111 stress-test patterns), Mask-OOD (500 unseen LithoBench masks), Process-OOD (extreme focus/dose) and Source-OOD (an unseen Quasar source). Besides PSNR/LPIPS/FID, a fine-tuned SAM extracts contours from both predictions and targets, from which Dice/B-IoU, gauge-based CD and EPE, and rule-based defect labels are computed. Seven baselines are compared (CNN, Transformer, two FNOs, Pix2Pix, two SD variants). The main message is that good FID or contour overlap can go together with large CD errors and poor defect prediction, especially under process shift. Model rankings are also checked against held-out wafer SEM images.

Critical strengths
  1. The metric design is the paper's most valuable contribution. Gauge-based CD/EPE and defect labels, computed with the same extractor on prediction and target, are much closer to how lithography engineers judge outputs than FID or PSNR, and they actually change the ranking: on Process-OOD, SD-Adapter has the best FID but a 55 nm CD error, against 9–11 nm for MFNO and CNN (Table 3).

  2. The OOD splits are sensible. Each one moves a single factor (mask, focus/dose, or source) and keeps the rest fixed, so it is clear what a failure means. Having any comparison against real wafer SEM is also more than most simulation-based benchmarks offer.

Critical weaknesses
  1. The fabrication metrics that carry the main finding do not behave as expected. EPE stays at 2–4.6 nm while CD error ranges from 9 to 113 nm (e.g., CNN on ID: 2.70 vs. 22.66), which is hard to explain unless a few gauges with bridges or pinches dominate the mean; in that case "good overlap, bad CD" is partly a defect effect rather than a linewidth one. The defect metric does not separate the models on Process-OOD either, where all seven score 0.34–0.36 Status F1 with overlapping CIs (Table 3).

  2. The real-wafer validation is thin. Table 4 gives no sample size or CIs, the rank correlations in Sec. 4.2 are computed over only six models, and the main text does not say which process conditions the wafer data cover. So it is unclear whether the Process-/Source-OOD conclusions, which carry the paper's main claims about model families, hold on real data.

Questions for authors
  1. How are gauges counted when the prediction has no valid edge pair, e.g., because of a bridge or a missing line? The median CD error, or CD on gauges valid in both prediction and target, would largely settle my first concern.

  2. How large is the wafer set in Table 4, and which focus/dose conditions and sources does it cover? If it includes off-nominal conditions, could you report Table 4 per condition with CIs, to show whether the Process-OOD finding (deterministic models keep CD error lower than SD-based ones) also holds on real data?

  3. For the Process-OOD defect results, could you add a trivial baseline, such as always predicting "defective" or a model that ignores F and E, to show whether any model does better than chance?

Minor: the Discussion paragraph repeats one sentence, and Yang et al. 2022a and 2022b are the same reference.

6 comments
AuthorsOpenReview ↗

Response to Reviewer SvEj (1/4)

Response to W1(1) and Q1(1–2): Large CD Errors and Gauge Evaluation

We thank the reviewer for the careful observation regarding the discrepancy between EPE and CD error. We agree that large mean CD errors may be influenced by localized geometric deviations caused by defects such as bridges and pinches. We clarify the gauge evaluation rules and provide additional analyses on the ID dataset to investigate this effect.

Gauge validity and CD calculation. Gauge validity is determined by whether a valid edge pair can be detected along the scanline. In this setting, a bridge may cause the detected edge pair to lie outside the original feature, resulting in large CD errors. We retain such gauges because their CD errors capture defect-induced geometric changes. Specifically, the CD error is calculated as follows:

GT GaugePrediction GaugeCD Error
ValidValid∣CDpred−CDGT∣\lvert CD_{\mathrm{pred}} - CD_{\mathrm{GT}} \rvert
ValidInvalidCDGTCD_{\mathrm{GT}}
InvalidValidExcluded
InvalidInvalidExcluded

This protocol preserves measurable defect-induced geometric errors while accounting for missing edge pairs in the predictions.

Additional CD error analysis. To examine the influence of large errors and localized defects, we report the median absolute CD error and the CD MAE after excluding cases affected by feature-disrupting defects, which displace the detected edge pair from the intended feature.


Additional CD error analysis on ID Dataset

ModelCD MAE w/o Feature-Disrupting Defects (nm)Median CD Error (nm)
CNN6.0852.584
Trans7.7692.201
RFNO10.5473.178
MFNO10.0383.009
Pix2Pix4.9661.278
SD-AdaIN10.8523.018
SD-Adapter5.2781.401

As shown, the median CD errors range from 1.278 to 3.178 nm, indicating that large errors in the distribution largely influence the mean. For CNN, the median CD error is only 2.584 nm, compared with its original mean CD error of 22.66 nm. After excluding feature-disrupting defects, its CD MAE decreases to 6.085 nm, a reduction of 73.1%. Across all seven models, the CD MAE after exclusion ranges from 4.966 to 10.852 nm. These results indicate that localized defect-induced geometric deviations contribute to large mean CD errors.

Interpretation of EPE and CD differences. EPE aggregates contour-edge displacement, while CD measures local dimensions using edge pairs along predefined gauge scanlines. Bridges or pinches can alter local topology and displace the measured edges, producing large CD errors even when the remaining contours maintain good agreement with the ground truth. This explains how CNN can exhibit an EPE of 2.70 nm alongside a mean CD error of 22.66 nm.

We agree that large CD errors can reflect both linewidth deviations and defect-related geometric changes. These deviations represent meaningful fabrication prediction errors. Accordingly, CD captures local dimensional variations and defect-related geometric changes at designated gauges, complementing overlap metrics that assess overall geometric agreement and Defect Status/Type metrics that evaluate defect occurrence and classification without directly quantifying geometric severity.

We have added these analyses to Appendix G.1 and further provide qualitative examples in Fig. 5 to illustrate the complementary nature of these metrics in the main paper.

AuthorsOpenReview ↗

Response to Reviewer SvEj (2/4)

Response to W1(2) and Q3: Defect Evaluation on Process-OOD

We thank the reviewer for the careful observation regarding the defect evaluation on Process-OOD. We agree that the Status F1 scores of the seven models are closely clustered between 0.340 and 0.356, with overlapping confidence intervals. These results do not provide sufficient evidence to distinguish the models in defect status prediction under Process-OOD conditions.

To address this concern, we evaluated two constant prediction policies on the same Process-OOD test set: Always-Defective, and Always-All-Types. The results are summarized below.


MethodStatus F1Type macro-F1Type micro-F1
CNN0.354 ± 0.0860.144 ± 0.0440.110 ± 0.035
Trans0.340 ± 0.0870.183 ± 0.0420.117 ± 0.034
RFNO0.354 ± 0.0850.209 ± 0.0390.161 ± 0.036
MFNO0.355 ± 0.0850.183 ± 0.0420.129 ± 0.035
Pix2Pix0.347 ± 0.0920.183 ± 0.0390.143 ± 0.036
SD-AdaIN0.353 ± 0.1200.222 ± 0.0680.138 ± 0.041
SD-Adapter0.356 ± 0.0870.235 ± 0.0370.135 ± 0.035
Always-Defective0.2759 ± 0.0244N/AN/A
Always-All-Types0.2759 ± 0.02440.0607 ± 0.00600.0620 ± 0.0062

We highlight two findings from these results:

  • (1) All seven models achieve higher Status F1 than the Always-Defective baseline (0.340–0.356 vs. 0.2759), suggesting that they capture defect-related information beyond a constant prediction strategy.

  • (2) Defect Type F1 reveals additional differences among models. On Process-OOD, where the defect distribution is imbalanced, the seven models exhibit closely clustered Status F1 scores (0.340–0.356), while their Type macro-F1 scores show greater variation (0.144–0.235). For example, CNN and SD-Adapter achieve nearly identical Status F1 scores (0.354 vs. 0.356), but different Type macro-F1 scores (0.144 vs. 0.235). This distinction arises because Status F1 evaluates defect occurrence without distinguishing defect categories, while Type F1 additionally evaluates classification across defect types.

These results demonstrate the complementary information provided by Status F1 and Type F1, particularly for challenging datasets with imbalanced defect distributions such as Process-OOD. In future work, we will further expand the coverage of layout structures and process conditions to improve the statistical representation of rare defect categories.

AuthorsOpenReview ↗

Response to Reviewer SvEj (3/4)

Response to W2(1) and Q2(1): Real-wafer Dataset Clarification and Statistical Results

We thank the reviewer for pointing out that the real-wafer validation in Table 4 (Table 5 in the revised version) lacked sufficient details. We have supplemented the dataset description in the main text.

Specifically, the real-wafer validation uses the test set described in Table 5 (Table 6 in the revised version), comprising approximately 2,000 SEM images across 200 masks. Regarding process-condition coverage, we would like to clarify that deviations from nominal process conditions do not necessarily imply Process-OOD conditions. The test set includes 9 focus levels spanning [-28 nm, 28 nm] and 11 dose levels spanning [20 mJ/cm2, 40 mJ/cm2], all acquired under quadrupole illumination. Although the test set includes off-nominal conditions, all evaluated conditions fall within the predefined Process-ID range.

To address the missing statistical information, we have updated Table 4 (Table 5 in the revised version) to include the corresponding 95% confidence intervals for each evaluation metric. The updated results are presented below.


Held-out Wafer-SEM Evaluation

ModelB-IoU (Mean)B-IoU (95% CI)EPE (Mean)EPE (95% CI)CD (Mean)CD (95% CI)Status F1 (Mean)Status F1 (95% CI)
CNN0.689±0.07613.19±0.563924.40±3.71930.771±0.0228
Transformer0.648±0.06583.76±0.562430.75±2.86900.674±0.0215
RFNO0.660±0.10643.91±0.431319.80±3.97130.802±0.0208
MFNO0.670±0.06973.38±0.590315.48±2.99610.818±0.0193
Pix2Pix0.682±0.08983.25±0.780114.68±3.27300.849±0.0201
SD-Adapter0.691±0.07493.12±0.820719.84±10.45550.867±0.0198

Furthermore, we have clarified the Rank Correlation analysis in Section 4.2. The Spearman correlation coefficients quantify the overall ranking consistency of the six evaluated models between FabBench and the real-wafer Process-ID evaluation. To provide a more transparent comparison beyond the correlation coefficients, we have also provided the individual rankings of all six models under both evaluation settings in Appendix G.2, Table 11. This allows readers to directly examine the correspondence between model rankings and identify any changes in their relative ordering.

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

This paper proposes FabBench, a fabrication-informed benchmark for process-conditioned lithography prediction in the SEM domain. The authors construct 547K mask–process–SEM triplets using physics-based lithography simulation and a real-SEM-trained emulator. The benchmark introduces structured ID/OOD splits and fabrication-oriented metrics, including CD, EPE, and defect prediction. Experiments across different model families reveal that high image similarity does not necessarily guarantee accurate geometric or defect predictions.

Critical strengths
  1. A large-scale benchmark covering continuous process variations and different mask patterns.
  2. The evaluation protocol incorporates fabrication-relevant metrics beyond conventional image similarity.
  3. Structured OOD splits enable systematic evaluation of model generalization.
  4. Extensive baseline comparisons and real-wafer validation demonstrate the potential value of the benchmark.
Critical weaknesses
  1. Existing datasets, such as LithoSim, already cover mask patterns and continuous process variations. Although FabBench introduces SEM-domain prediction and defect evaluation, the paper lacks direct comparisons demonstrating the additional benefits over existing benchmarks.
  2. Defect prediction is a major motivation of this work. However, some defect categories contain very few samples, especially in OOD splits, which may affect the reliability of the reported results.
  3. The geometric and defect metrics rely on SAM-based contour extraction and rule-based defect labeling. Errors in these intermediate steps may affect the evaluation results and model rankings.
Questions for authors
  1. Can the authors provide a more direct comparison with LithoSim or other existing benchmarks to demonstrate the unique benefits of FabBench?
  2. How sensitive are the evaluation results to errors introduced by SAM-based contour extraction and defect labeling?
2 comments
AuthorsOpenReview ↗

Response to Reviewer goGd (1/2)

Response to W1 and Q1: Comparison with LithoSim and Existing Benchmarks

We thank the reviewer for pointing out the need for a more direct comparison between FabBench and existing lithography benchmarks. To this end, we compare FabBench with existing benchmarks from multiple perspectives, as summarized in the following table.


BenchmarkPrediction TargetSEM Image FidelityGlobal Geometry EvaluationSEM-domain Local MetrologyTopological Defect EvaluationGeneralization Evaluation
CAD13Resist—Resist geometry——Standard evaluation
ISPD19Resist—Resist geometry——Standard evaluation
LithoBenchResist—MSE / IoU / PA——Layout generalization
LMLithoResist—mIoU / mPA / EPE——Cross-node transfer
LithoSimResist—IoU / EPE——Joint OOD evaluation
FabBenchSEMPSNR / LPIPS / FIDDice / B-IoU / EPECDStatus / Type F1Mask / Process / Source-OOD

--

This comparison highlights the additional evaluation capabilities of FabBench. Existing benchmarks primarily focus on geometric prediction and evaluation in the simulated resist domain, whereas FabBench extends the evaluation to the SEM observation domain, integrating image appearance, global contour geometry, local metrology, and topological defect evaluation into a hierarchical assessment framework, from visual fidelity to fabrication-relevant accuracy. Furthermore, by introducing separate Mask-OOD, Process-OOD, and Source-OOD splits, FabBench enables the evaluation of model performance under distinct types of distribution shifts.

In addition, we have provided a quantitative comparison with LithoSim in Appendix G.5 (Table 16). Taking Transformer and MFNO as examples, their contour IoU scores on LithoSim are 91.69 and 92.18, respectively, indicating comparable global geometric agreement. However, their CD errors on FabBench differ substantially, at 41.55 nm and 19.05 nm, respectively. This demonstrates that similar global contour overlap can conceal substantial differences in local dimensional accuracy, highlighting the importance of incorporating local metrology into lithography model evaluation. Given that the two benchmarks use different datasets, we present this result as complementary cross-benchmark evidence, rather than a strictly controlled ranking comparison on the same dataset.

Through the additional benchmark comparison table and the quantitative analysis above, we aim to clarify the expanded evaluation scope and practical value of FabBench compared with existing benchmarks.

AuthorsOpenReview ↗

Response to Reviewer goGd (2/2)

Response to W2: Rare Defect Classes and Evaluation Uncertainty.

To provide a more comprehensive assessment under such class imbalance, we report both macro-F1 and micro-F1 for defect-type prediction in Tables 3–4 with CIs. Macro-F1 assigns equal weight to each defect category, whereas micro-F1 reflects overall prediction performance by aggregating predictions across categories. Appendix G.3 (Table 13) further details the exact number of positive samples for each defect category across all splits.

For Process-OOD, we acknowledge that the Status F1 scores of the seven models are closely clustered, providing insufficient evidence to distinguish their defect status prediction performance. To further contextualize these results, we compare the seven models against two constant prediction policies, Always-Defective and Always-All-Types, on the same test set:


MethodStatus F1Type macro-F1Type micro-F1
CNN0.354 ± 0.0860.144 ± 0.0440.110 ± 0.035
Trans0.340 ± 0.0870.183 ± 0.0420.117 ± 0.034
RFNO0.354 ± 0.0850.209 ± 0.0390.161 ± 0.036
MFNO0.355 ± 0.0850.183 ± 0.0420.129 ± 0.035
Pix2Pix0.347 ± 0.0920.183 ± 0.0390.143 ± 0.036
SD-AdaIN0.353 ± 0.1200.222 ± 0.0680.138 ± 0.041
SD-Adapter0.356 ± 0.0870.235 ± 0.0370.135 ± 0.035
Always-Defective0.2759 ± 0.0244N/AN/A
Always-All-Types0.2759 ± 0.02440.0607 ± 0.00600.0620 ± 0.0062

We highlight two findings from these results:

  • (1) All seven models achieve higher Status F1 than the Always-Defective baseline (0.340–0.356 vs. 0.2759), suggesting that they capture defect-related information beyond a constant prediction strategy.

  • (2) Defect Type F1 reveals additional differences among models. On Process-OOD, where the defect distribution is imbalanced, the seven models exhibit closely clustered Status F1 scores (0.340–0.356), while their Type macro-F1 scores show greater variation (0.144–0.235). For example, CNN and SD-Adapter achieve nearly identical Status F1 scores (0.354 vs. 0.356), but different Type macro-F1 scores (0.144 vs. 0.235). This distinction arises because Status F1 evaluates defect occurrence without distinguishing defect categories, while Type F1 additionally evaluates classification across defect types.

These results demonstrate the complementary information provided by Status F1 and Type F1 for challenging datasets. We are now actively expanding the coverage of layout structures and process conditions to improve the representation of rare defect categories and strengthen the statistical robustness of the evaluation.

Response to W3 and Q2: Sensitivity to SAM Contour Extraction and Defect Labeling.

To assess the reliability of the intermediate steps in our evaluation pipeline, we conducted sensitivity analyses of contour extraction and defect labeling to examine their potential impact on the final evaluation results

  • SAM-based contour extraction: To assess the impact of SAM contour extraction errors on geometric metrics and model comparisons, we conducted an independent expert evaluation on 200 test samples. We selected one representative model from each of the three model families: Transformer, MFNO, and SD-Adapter. For the predictions generated by these models and their corresponding ground-truth SEM images, we obtained top and bottom contours using both SAM-based automatic extraction and expert annotation. The results are summarized below.

    ModelCD (SAM) ↓CD (Expert) ↓EPE (SAM) ↓EPE (Expert) ↓B-IoU (SAM) ↑B-IoU (Expert) ↑
    Transformer20.8219.892.632.480.6210.617
    MFNO18.1817.622.212.350.6500.658
    SD-Adapter22.7421.021.942.160.6970.684

Although the two evaluation protocols yield some differences in the metric values, the relative rankings of the three models remains across all three metrics. Specifically, the CD ranking is MFNO, Trans, and SD-Adapter, while the EPE and B-IoU rankings are both SD-Adapter, MFNO, and Trans. These results indicate that, for the representative models and test samples evaluated, SAM-based contour extraction errors do not alter the relative performance rankings.

  • Rule-based defect labeling: For rule-based defect labeling, we have conducted a threshold sensitivity analysis in Appendix G.4. Specifically, Missing is the only defect category that involves tunable thresholds. We evaluated 27 combinations of its three decision thresholds and measured the resulting label-change rates and Jaccard similarities. The maximum label-change rates across the ID, Mask-OOD, Process-OOD, and Source-OOD splits are 0.499%, 0.044%, 0.265%, and 0.901%, respectively. These results indicate that the defect annotations are largely stable under the tested threshold variations, with fewer than 1% of labels changing across all four splits
Reviewer_RRQUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FabBench is a benchmark for process-conditioned lithography models that predict SEM-domain images from a mask layout and continuous process conditions, namely illumination source, defocus, and dose. The authors identify three gaps (1) existing benchmarks either focus purely on physics simulation or isolate defect analysis from SEM-domain observations (2) the scarcity of process-annotated wafer data prevents joint evaluation of geometric evolution and defect formation and (3) conventional image-level metrics are poorly aligned with fabrication-side assessment. To address these, FabBench couples a rigorous lithography simulator with a generative SEM emulator trained on real SEM data, producing 547K physics-grounded, real-SEM-informed mask-process-SEM triplets with ID, Mask-OOD, Process-OOD, and Source-OOD splits. Its distinctive component is a hierarchical, metrology-aware evaluation protocol spanning visual fidelity, contour geometry, gauge-based CD and EPE, and topological defect consistency. The authors show that generative models achieve higher visual fidelity while deterministic and operator-based models exhibit more stable local metrology, and that high contour overlap can still coincide with substantial CD errors and incorrect defect states, a trend they report to be consistent on held-out wafer SEM.

Critical strengths
  1. A clear and relevant motivation. The paper shows that similar image-overlap scores can hide large differences in local measurement accuracy. This makes a convincing case for evaluating lithography models beyond standard image metrics.

  2. Useful OOD evaluation. Separating changes in masks, process conditions, and illumination makes it easier to understand where models fail to generalize.

  3. Acknowledged limitations and real-wafer checks. The paper is clear about the limitations of synthetic SEM targets and includes real-wafer evaluations, including manually measured CD values independent of the automated evaluator.

Critical weaknesses
  1. The benefit of SEM rendering is not fully established. The emulator adds realistic appearance, but the underlying geometry still comes from simulation. A comparison with direct contour prediction would help show whether rendering improves fabrication-relevant evaluation and how much error it introduces.

  2. The evaluation may depend on image style. All geometric and defect scores rely on the same learned contour extractor. Its overall accuracy does not establish that it measures outputs from different model families equally well. An independent evaluation of a representative subset would make the comparisons more convincing.

  3. The model-family conclusions need more support. The differences between PSNR and perceptual metrics resemble the familiar perception–distortion trade-off. The metrology results are interesting, but architecture, conditioning, and training objectives also differ across methods. It is therefore difficult to attribute the observed behavior to model family alone.

  4. Several relevant methods are missing. Recent lithography models such as Unitho, LMLitho, and LithoDreamer are discussed but not evaluated. Including an appropriately adapted representative, or explaining their exclusion, would strengthen the benchmark’s coverage.

  5. Some generalization claims rest on limited coverage. Despite the large sample count, training uses a relatively small set of masks, and the process and illumination tests cover a narrow range of shifts. Some defect categories also have very few positive examples. Clearer split reporting and more cautious interpretation of these results are needed.

Questions for authors

Please see the weaknesses section.

9 comments
AuthorsOpenReview ↗

Response to Reviewer RRQU (1/3)

Response to W1: Role and Validation of SEM Rendering in FabBench..

We thank the reviewer for the comment and would like to clarify the role of SEM Rendering. SEM rendering is not intended to improve simulated geometric accuracy, but to establish the SEM-domain prediction task evaluated by FabBench. Acquiring real SEM data across a large number of Mask–Process combinations requires extensive wafer exposure and metrology experiments, which are costly and difficult to scale. To address this challenge, we employ physics-based lithography simulation to generate physically grounded resist geometries under different mask layouts and process conditions. An SEM emulator trained on real SEM data then transforms these simulated geometries into SEM-domain targets, incorporating realistic observation characteristics such as edge enhancement and stochastic textures. This enables FabBench to evaluate models that predict SEM images directly from mask layouts and process conditions.

Direct contour prediction cannot replace SEM prediction because the two tasks have different prediction targets. Direct contour prediction maps mask layouts and process conditions to resist contours, whereas FabBench evaluates SEM images. This distinction is particularly important for real-wafer validation, where the actual observations are CD-SEM images rather than idealized contours. Evaluating contours alone assesses geometric agreement but cannot capture SEM-specific observation characteristics, which are particularly relevant in advanced-node manufacturing.

To quantify the error introduced by SEM rendering, we have conducted a dedicated component validation using independently held-out real-wafer SEM data in Appendix C. Specifically, contours extracted from real SEM images were provided as inputs to the emulator. By using experimentally observed contours rather than simulation-generated geometries, this validation isolates rendering fidelity from the accuracy of the simulated geometry. The rendered SEM images were compared with their corresponding real SEM images, achieving a CD error of 2.12 nm, an EPE of 1.24 nm, a Dice score of 98.84%, a B-IoU of 92.17%, and a Defect Type F1 of 93.10%. These results quantify the discrepancy introduced by the rendering stage and demonstrate its ability to reproduce real SEM observations while preserving geometric metrology information and topological defect characteristics under this validation setting.

Response to W2: Rely on the Same Contour Extractor and Independent Validation of SAM-Based Evaluation.

We thank the reviewer for raising concerns about evaluation fairness across different model families. FabBench aims to evaluate models' ability to generate SEM images, with a unified contour extractor used to assess geometric and defect characteristics from the generated outputs. We apply the same fixed SAM-based contour extractor to all model predictions to ensure a consistent evaluation protocol. To further assess the consistency of this evaluation protocol across different model families, we conducted an independent manual contour evaluation.

We test 200 samples and one representative model from each of the three model families: Transformer (deterministic prediction), MFNO (neural operator learning), and SD-Adapter (conditional generation). For the predicted SEM images from these models and their corresponding ground-truth SEM images, three annotators independently annotated the contours and the resulting metric values were averaged across annotators. We also extracted contours from the same images using SAM for comparison.

ModelCD (SAM) ↓CD (Expert) ↓EPE (SAM) ↓EPE (Expert) ↓B-IoU (SAM) ↑B-IoU (Expert) ↑
Transformer20.8219.892.632.480.6210.617
MFNO18.1817.622.212.350.6500.658
SD-Adapter22.7421.021.942.160.6970.684

The results show that the metric differences between SAM-based and manual evaluations are small, with no evident model-specific bias large enough to alter the relative rankings among the three evaluated models. This suggests that the observed performance differences largely reflect differences in their generated SEM outputs. Due to the time constraints of the rebuttal period, we limited this analysis to three representative models and 200 test samples. We will extend the evaluation to all benchmark models and a larger set of test samples in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer RRQU (2/3)

Response to W3: Interpretation of Model-Family Results.

We thank the reviewer for the thoughtful comment. We would like to clarify that FabBench categorizes model families based on their architectural and modeling paradigms (Sec. 4.1), and evaluates representative baselines using the same training data and evaluation protocol, while retaining their standard training objectives. Our conclusions in Sec. 4.2 therefore describe empirical trends across the evaluated baselines, rather than attributing the observed differences solely to model family. Moreover, we explicitly investigate variations within the same family: Appendix H.3 and Tables 3–4 compare different conditioning mechanisms for Stable Diffusion, demonstrating that conditioning design also affects metrology performance.

Response to W4: Baseline Selection and LMLitho Adaptation.

We thank the reviewer for pointing out the absence of several relevant methods in our benchmark comparison. We address this concern from two aspects:

  • Task compatibility and baseline selection. Unitho and LithoDreamer involve upstream mask generation and multi-stage lithography modeling, respectively, whereas FabBench takes a fixed mask as input and evaluates process-conditioned SEM prediction. LMLitho is more closely related, as it predicts resist images from mask designs. However, its original formulation does not cover the full process-conditioning and SEM-domain prediction setting of FabBench. Furthermore, complete official implementations of these methods were not publicly available at the time of our rebuttal, making direct reproduction difficult. We therefore implemented an adapted LMLitho baseline (LMLitho-adapted), introducing the conditioning mechanism used in Trans to align it with the FabBench task.

  • Additional evaluation of LMLitho-adapted. We evaluated LMLitho-adapted on all four FabBench splits using the same evaluation protocol. The results below reveal a consistent difference between image distribution similarity and fabrication-related accuracy. LMLitho-adapted achieves lower FID than CNN across all four splits, indicating better distributional similarity. However, CNN consistently achieves better B-IoU, EPE, and CD across all four splits, as well as higher defect-status F1. These results demonstrate that improved image distribution similarity does not necessarily translate into better local geometric accuracy or defect prediction, highlighting the additional evaluation value provided by FabBench.

Test SplitMethodPSNR ↑ (dB)LPIPS ↓FID ↓Dice ↑B-IoU ↑EPE ↓ (nm)CD ↓ (nm)Status F1 ↑Type Macro-F1 ↑Type Micro-F1 ↑TAT ↓
IDCNN24.4390.658163.170.9560.7632.7022.660.7090.3980.4180.029
IDLMLitho-adapted23.8810.667159.4300.9300.7283.39535.590.7020.3730.4010.07
Mask-OODCNN23.9610.718144.640.9540.7371.9824.220.8370.4160.6120.029
Mask-OODLMLitho-adapted23.6060.727137.6140.9610.7152.09924.850.8230.3620.4970.071
Process-OODCNN24.2960.721243.4100.922 ± 0.0230.771 ± 0.0423.08 ± 0.7810.78 ± 8.150.354 ± 0.0860.144 ± 0.0440.110 ± 0.0350.029
Process-OODLMLitho-adapted24.2020.718240.4030.915 ± 0.00630.725 ± 0.01323.98 ± 0.2618.90 ± 10.070.318 ± 0.02910.161 ± 0.01810.1106 ± 0.01850.071
Source-OODCNN19.7670.631181.4280.927 ± 0.0140.627 ± 0.0442.94 ± 0.5222.66 ± 11.860.451 ± 0.1430.121 ± 0.0400.137 ± 0.0380.029
Source-OODLMLitho-adapted19.0660.627154.9960.899 ± 0.00690.602 ± 0.01553.66 ± 0.2331.27 ± 12.060.391 ± 0.04860.089 ± 0.01830.126 ± 0.02670.071

We have added LMLitho-adapted to the main benchmark results for all four splits in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer RRQU (3/3)

Response to W5: Clearer Reporting Split Reporting and More Cautious Interpretation.

We thank the reviewer for highlighting the importance of data coverage in interpreting generalization results. We agree that generalization claims should be considered within the scope of the evaluated distribution shifts. The rationale and scope of the OOD splits are detailed in Sec. 3.3. Specifically:

  • Mask generalization. While the training set contains 111 masks, FabBench evaluates structural generalization on 500 unseen masks (Mask-OOD). This provides a dedicated assessment of generalization beyond the training layouts.
  • Process generalization. Process-OOD covers 17 selected near-boundary conditions to evaluate extrapolation under extreme process variations while avoiding complete pattern failure. The findings characterize model behavior under these challenging process variations.
  • Source generalization. Source-OOD evaluates an unseen Quasar illumination family under controlled conditions. The findings are therefore specific to this illumination-family shift.
  • Defect evaluation and interpretation. We report macro-F1 and micro-F1 (Tables 3–4), per-class positive sample counts (Appendix G.4, Table 14), and bootstrap 95% confidence intervals (Table 4). Accordingly, we have interpreted results for rare defect categories and small inter-model differences, particularly under Process-OOD.

Overall, these clarifications explicitly define the coverage of each OOD split and the corresponding scope of our generalization findings.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank the reviewers and the AC for their time and careful reading. Reviewers highlighted the value of fabrication-oriented metrics (RRQU, goGd, SvEj), the structured OOD splits for evaluation (RRQU, goGd, SvEj), and the inclusion of wafer-SEM evaluations (RRQU, goGd, SvEj).

Scope of revision. In the main text, we add a comparison with existing lithography benchmarks, extend the evaluation of LMLitho-adapted to all four FabBench splits, and clarify the complementary roles of contour overlap, local metrology, and defect prediction in capturing distinct fabrication-related failure modes. In the appendix, we clarify the gauge-based CD calculation procedure and provide additional analyses to further support the reliability of the evaluation. Key textual 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.