acceptodds
Under review as a conference paper at ICLR 2027

Flow-Litho: A Physics-Guided World Model for Faithful Process Responses in Multi-Stage Computational Lithography

Abstract

Computational lithography improves pattern fidelity and manufacturability in advanced integrated circuit (IC) fabrication through modeling and optimization of pattern transfer. However, process responses vary substantially across stages and local regions, making reliable multi-stage lithography modeling challenging. To this end, we propose FlowLitho, a physics-guided lithography world model (WM) for faithful process-response modeling, which recursively evolves the "Layout-Mask-Resist Image-ADI" pipeline through intra-stage physics correction and cross-stage difference preservation. First, we formulate the heterogeneous multi-stage process as a latent-space flow-matching problem, learning process-conditioned velocity fields and evolving latent states through numerical integration. Second, we develop a physical dynamics correction mechanism that evaluates regional response discrepancies using stage-specific approximate physical operators and couples their gradients to correct the latent velocity field. Finally, we introduce a difference-preserving self-evolving optimization paradigm for joint training, which uses paired supervision for the same layout across process conditions to guide intermediate-state recomposition and downstream replay, enabling the model to preserve process-induced differences while adapting to upstream rollout deviations. Experiments show that Flow-Litho achieves state-of-the-art performance in multi-stage lithography modeling and process-parameter planning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
World Model, Lithography, Image Process

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_G5X3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Flow-Litho, a physics-guided world model for multi-stage computational lithography. It combines latent-space flow matching, physics-guided velocity correction, and self-evolving optimization to improve lithography prediction and process-parameter planning. Experiments demonstrate improved accuracy compared with existing methods.

Critical strengths
  1. A unified framework for multi-stage lithography modeling using flow matching.
  2. Physics-guided correction incorporates physical information into the prediction process.
  3. Extensive experiments and ablation studies demonstrate promising improvements over existing methods.
Critical weaknesses
  1. LithoDreamer has already introduced world models for multi-stage lithography. The proposed method mainly combines existing flow-matching and physics-guided techniques, with limited discussion of its fundamental novelty.
  2. Section 3.2 divides images into four fixed quadrants, which may not capture highly localized lithography hotspots. The choice lacks sufficient justification.
  3. Only one unseen 55 nm process condition is evaluated, which is insufficient to support broad claims of process robustness.
  4. Physics-guided correction requires additional gradient computations at each step, but the computational overhead is not adequately evaluated.
Questions for authors
  1. Why use four fixed quadrants instead of hotspot-aware regions?
  2. What is the inference-time overhead of physics-guided correction?
1 comment
AuthorsOpenReview ↗

Clarification and Discussion of the Problem

Thank you for your suggestions. We respond to your comments below.

W1: Novelty Beyond LithoDreamer.

A: Flow-Litho's novelty is not a new world model or flow-matching method, but its treatment of physical-response constraints and cross-stage error propagation. (1) Physical representation: LithoDreamer uses stage-specific latent constraints and intervention policies. Flow-Litho embeds lithography-equation-based approximate operators into latent dynamics and corrects latent velocity through regional response discrepancies and coupled gradients at each integration step (Secs. 2.1, 3.2). (2) Cross-stage optimization: Difference-preserving self-evolving optimization uses same-layout, cross-condition supervision for intermediate-state recomposition, followed by downstream replay to adapt to upstream rollout deviations (Sec. 3.3). (3) Evidence: Versus LithoDreamer, Flow-Litho gains 6.81/4.94 pp in ADI mIoU on 55 nm ID/unseen-condition tests, with ADI EPEs of 3.67/3.93 nm; on 28 nm, Resist Image EPE is 43.8% lower. Under 12-rollout GP-BO, Accp rises from 55.57% to 78.29%, and target ADI EPE falls from 5.65 to 4.08 nm (Tabs. 1, 3). Removing physical correction or self-evolving optimization reduces Accp by 31.66/31.90 pp (Tab. 4). Thus, the contribution is integrating equation-based regional correction and cross-condition difference preservation into multi-stage latent dynamics.

W2,Q1: About Quadrant.

A: Four fixed quadrants provide deterministic, non-overlapping supports over all valid pixels without requiring a separate hotspot detector. A morphological boundary band is computed on the full image before partitioning to emphasize localized contour deviations; the region-wide term captures broader structural errors (Sec. 3.2.1; App. E.1).

Tab. C.6 shows that four-region coordination outperforms one global region: ADI EPE falls from 4.73 to 3.67 nm and Accp rises from 54.99% to 78.29%. Boundary-only and region-only variants achieve 62.49% and 68.64% Accp.

Direct comparison: With four regions and the same correction metrics and gradient-coupling solve, hotspot-aware partitioning achieves 84.92% ADI mIoU, 4.06 nm EPE, and 72.47% Accp, versus 85.08%, 3.67 nm, and 78.29% for fixed quadrants (Tab. R.1). Fixed quadrants yield comparable mIoU, 0.39 nm lower EPE, and 5.82 pp higher Accp.

Table R.1. Fixed vs. Hotspot-aware Partitioning

StrategyNumber of RegionsADI mIoUADI EPEExact Recovery
Geometry-aware Hotspot Partitioning484.924.0672.47
Four quadrants (Ours)485.083.6778.29

W3: Generalization Scope.

A: The evidence covers three distinct settings. (1) Unseen layouts: The 55 nm ID test layouts are disjoint from training; both splits cover the same 36 observed configurations, so this tests layout generalization under seen conditions. (2) Unseen process conditions: We further evaluate an unseen joint process configuration (Annular, dose 1.0, defocus 0 nm, and threshold 0.119340), which is absent from training, demonstrating generalization to previously unseen process conditions. Flow-Litho achieves 80.98% ADI mIoU and 3.93 nm ADI EPE. (3) Cross-node generalization: More importantly, we extend evaluation from 55 nm to the independent 28 nm LithoSim open-source dataset, involving a substantially different technology node, unseen layout patterns, and node-specific process conditions, without using any 28 nm training data. This tests whether the learned process-response dynamics transfer across both layout and technology domains. Flow-Litho achieves 1.59 nm Mask EPE and 0.77 nm Resist Image EPE, with the latter 43.8% lower than LithoDreamer (Tab.1). Together, these results demonstrate generalization beyond a single unseen process condition, extending from unseen layouts and process conditions to cross-node transfer.

W4,Q2: Overhead.

A: At https://anonymous.4open.science/r/litho/Overhead.md: On one A100 80GB (batch 1, 512×512; FP16 network/FP32 physical computation), after 50 warm-ups and over 100 runs, the same checkpoint and nine Euler updates take 207 ms without inference correction and 304 ms with correction: +97 ms (46.9%) and +4.1 GB peak memory (8.2 to 12.3 GB) (Tab. O1). Each update computes four regional gradients and solves a 4×4 system. For 12-candidate GP-BO, Flow-Litho (PC-On) takes 3.79 s (3.648 s rollouts + 0.142 s search), versus 4.25 s for LithoDreamer; peak planning memory is 13.02 versus 15.32 GB (Tab. O3). PC-Off uses the same PC-trained checkpoint, with correction disabled only at inference. Relative to PC-Off, correction adds 1.11 s (2.68 to 3.79 s), while Accp rises from 58.49% to 78.29% and target ADI EPE falls from 5.84 to 4.08 nm. Thus, correction improves planning accuracy while retaining lower end-to-end latency and peak memory than LithoDreamer.

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

The paper proposes Flow-Litho for process-conditioned prediction of the Layout-Mask-Resist Image-After Development Image pipeline. A shared, stage-conditioned transformer learns latent velocity fields with three Euler updates per stage. Approximate physical operators define regional response discrepancies, and a regularized gradient-coupling system corrects the velocity. During joint training, predictions under two process conditions for the same layout are exchanged where their labels agree, followed by downstream replay. The model is evaluated on multi-stage prediction, discrete process-parameter recovery, and fixed-process trajectory reconstruction. The reported results show improvements over the compared methods, with several controlled ablations supporting the correction and replay mechanisms.

Critical strengths
  1. The method addresses a relevant difficulty: preserving small process-dependent changes while limiting error propagation across heterogeneous stages. Regional physical correction and difference-preserving replay are connected directly to this objective.

  2. The ablations go beyond removing entire modules. Tables 4 and C.4-C.6 compare gradient coupling, difference-preserving exchange, physical-loss supervision, replay gradients, and regional discrepancy construction under specified controls. These experiments provide useful evidence about the proposed mechanisms.

  3. The evaluation distinguishes exact parameter recovery from target-image agreement and within-model process-response agreement. The latter is not treated as a substitute for correct parameter identification. Several reported gains are substantial, including 78.29% versus 55.57% recovery against LithoDreamer under the same 12-candidate GP-BO budget. The separate 28 nm experiment also examines cross-node transfer.

Critical weaknesses
  1. The reported metrics are inconsistent with the stated definitions. For LMLitho on 55 nm ID Resist Image prediction, Table 1 reports mIoU=96.35%\mathrm{mIoU} = 96.35\%, whereas Table C.1 reports mPA=95.23%\mathrm{mPA} = 95.23\%. Appendix B explicitly states mIoU≤mPA\mathrm{mIoU} \leq \mathrm{mPA} under the shared binary protocol. There are further inconsistencies: those binary definitions imply that mPA−mIoU\mathrm{mPA} - \mathrm{mIoU} cannot exceed 50 percentage points, yet 28 nm LithoNet reports 93.06%93.06\% versus 31.49%31.49\%. Please reconcile the definitions, implementations, aggregation, and affected tables. Separately, reported EPE excludes unmatched gauges under Eq. (B.5), but matching-completeness fractions are not provided. The unmatched-gauge penalty used for planning does not apply to the reported EPE.

  2. Independence of layouts in the main data split is not established. The dataset includes 2,000 layouts repeated across all 36 process settings, but the 260,000/20,000 training/test split is described only in samples. Appendix A.2 establishes disjoint layout-target pairs, which does not necessarily imply disjoint layouts. Please specify the split unit, unique layout counts, and allocation of repeated conditions and any overlapping crops, including the paired subset. Appendix D.3 documents layout grouping for operator calibration, but does not resolve the main split. The unseen 55 nm threshold lies inside the training threshold range, so that experiment establishes interpolation to an unobserved condition rather than extrapolation beyond the process range. This should be distinguished from the separate cross-node result.

  3. Baseline adaptation and statistical reporting need clarification. The conditioning, architectural adaptation, training data, and training budgets of the generic world-model baselines are not specified comparably to Flow-Litho's own settings. Describing them as lithography-adapted does not explain how the comparison was implemented. These details affect attribution of the gains to the proposed mechanisms. Appendix A.2 also specifies layout-bootstrap confidence intervals, yet none accompany the planning results. Please provide the existing baseline protocols and the stated uncertainty estimates for the principal comparisons.

  4. The practical cost of physical correction is unmeasured. Corrected rollouts require physical-response evaluations, decoder differentiation, and regional gradients. The small coefficient solve and absence of additional trainable parameters do not establish low runtime. The 12-versus-36 comparison demonstrates fewer candidate rollouts within Flow-Litho, but does not establish lower end-to-end planning time relative to another model or simulator. Measured runtime and memory for the existing forward and planning settings, including correction-enabled and correction-disabled configurations on matched hardware, would make the claimed practical tradeoff assessable.

Questions for authors
  1. For the 55 nm ID resist task, Table 1 reports LMLitho mIoU of 96.35%, whereas Table C.1 reports mPA of 95.23%. How should these values be reconciled with the definitions in Appendix B, which imply mIoU <= mPA? Were both tables evaluated using the same implementation and aggregation?

  2. What fraction of target gauges is unmatched for each method in the forward-evolution experiments? Would the principal contour comparisons change if unmatched gauges were included using the penalty already defined for planning?

  3. For the main data split, are all process-condition instances of the same layout assigned to a single split? Please provide the unique layout counts in each split and explain how the 72,000 paired sequences and any overlapping crops are partitioned.

  4. How were the generic world-model baselines adapted to accept lithography process conditions, and what training data and budgets were used? Could you also provide the layout-bootstrap confidence intervals specified in Appendix A.2 for the main process-recovery comparison?

  5. What are the measured latencies of a complete forward rollout and the 12-candidate planning procedure, including physical correction and search overhead? How much of this cost is attributable to physical correction under matched hardware and inference settings?

1 comment
AuthorsOpenReview ↗

Clarification and Discussion of the Problem

Thank you for your suggestions. We respond to your comments below.

W1,Q1,Q2: Metrics and Gauge.

A: (1) LMLitho's 55-nm ID Resist Image values contained a copy-and-paste error; the corrected values are 95.23% mIoU and 96.35% mPA. During baseline reproduction, undefined mPA class terms (zero denominators) were propagated as image-level NaNs. Thus, affected images were excluded from mPA averaging but retained in mIoU averaging when defined, so the metrics used different image sets. We corrected the implementation to handle undefined class terms locally rather than propagating them as image-level NaNs. Both metrics are now averaged over the same valid images. An audit across all methods, stages, and conditions found this issue only in LithoNet, LMLitho, and Unitho. We recomputed the affected entries and updated Tabs.1,C.1; other methods were unaffected. (2) Our EPE_avg calculation method follows the commercial sEPE protocol, whereas EPE_plan is used for candidate selection. On 55 nm ID Mask, Flow-Litho, LithoDreamer, and DreamerV3 achieve 100% gauge matching. Unmatched-gauge fractions for DINO-WM, DIAMOND, OptiCo, Unitho, LMLitho, and LithoNet are 2.4%, 7.8%, 6.0%, 6.3%, 6.6%, and 8.3%, respectively. Applying the image-diagonal penalty used in planning gives penalty-adjusted EPEs of 225.08, 683.54, 517.36, 542.72, 566.68, and 746.61 nm. Although some baseline rankings change, Flow-Litho retains the lowest EPE (0.42 nm), confirming that its contour advantage is not an artifact of excluding unmatched gauges.

W2,Q3: Data.

A: We revise at https://anonymous.4open.science/r/litho/Data_Split.md: (1) Split: Layout ID is the unit: 1,845 training and 155 ID-test layouts, with no overlap. All 36 conditions and derived images per layout stay in one split. The 72,000 Layout–Condition sequences split into 66,420/5,580 (train/test), corresponding to 265,680/22,320 stage-specific image instances. (2) Pairs/crops: Cross-condition pairs, paired-subset examples, and overlapping crops inherit their Layout ID's split; none are independently randomized or cross splits. (3) OOD: The 3,000-sample 55-nm set uses Annular/dose 1.0/defocus 0 nm/unseen threshold 0.119340. This specific joint process configuration is unseen during training, with a threshold value absent from the observed settings; 28-nm LithoSim separately evaluates cross-node generalization. (4) Calibration: App.D.3 uses 50 training layouts for calibration and 20 disjoint validation layouts; this subdivision does not redefine the main split. We have corrected the dataset statistics and clarified the splitting protocol in the revised manuscript.

W3,Q4: About Baseline.

A: (1) Baseline adaptation. All baselines use the same 55 nm training data and split, retaining their respective architectures and training schedules. For generic world models, Action/Control inputs are replaced by process conditions p, held fixed throughout the three-stage rollout, without intermediate-action supervision. Checkpoints are selected on the same held-out validation set. (2) Statistical reliability. Following App.A.2, we perform 1,000 layout-bootstrap resamples, retaining all associated queries. Results are reported with 95% confidence intervals:

ModelExact Parameter Recovery (%)95% Layout Bootstrap CI
DINO-WM42.53[38.82, 46.21]
LithoDreamer55.57[52.91, 58.26]
Ours78.29[76.43, 80.12]

Flow-Litho improves recovery by 35.76 and 22.72 percentage points over DINO-WM and LithoDreamer, respectively. The non-overlapping confidence intervals support this separation across layout resamples.

W4,Q5: Computational Cost.

A: At https://anonymous.4open.science/r/litho/Overhead.md: (1) On one A100 80GB (batch size 1, 512×512), using the same checkpoint and nine Euler updates, physical correction increases complete rollout latency from 207 to 304 ms (+97 ms, 46.9%) and peak memory from 8.2 to 12.3 GB (+4.1 GB). PC-Off uses the same checkpoint trained with correction but disables correction at inference; PC-On retains inference-time correction. Each update computes four regional gradients and solves a 4×4 system (nine corrections per rollout) (Tab.O1). (2) Under matched hardware and input settings, Flow-Litho (PC-On) takes 304 ms/rollout, versus 323 ms for LithoDreamer, while improving ADI mIoU by 6.81 percentage points and reducing EPE from 3.74 to 3.67 nm (Tab.O2). (3) Under the same 12-candidate GP-BO protocol, PC-Off takes 2.68 s (2.484 s rollout + 0.196 s search overhead), PC-On takes 3.79 s (3.648 s rollout + 0.142 s search overhead), and LithoDreamer takes 4.25 s (3.876 s rollout + 0.374 s search overhead). Relative to PC-Off, PC-On adds 1.11 s but raises recovery from 58.49% to 78.29% and reduces target EPE from 5.84 to 4.08 nm. Peak planning memory is 13.02 GB for PC-On, 8.79 GB for PC-Off, and 15.32 GB for LithoDreamer.

Reviewer_1SqiOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Flow-Litho is a physics-guided lithography world model for faithful process-response modeling. It recursively evolves the Layout-Mask-Resist Image-After Development Image (ADI) pipeline under relative dose, defocus, effective resist threshold, and source category. The authors identify three challenges (1) stage heterogeneity (2) local non-uniformity, and (3) cross-stage coupling. For stage heterogeneity, a shared DiT learns process-conditioned velocity fields through latent-space flow matching. For local non-uniformity, a physical dynamics correction mechanism evaluates regional response discrepancies using stage-specific approximate physical operators and couples their gradients to correct the latent velocity field, without introducing additional trainable parameters. For cross-stage coupling, a difference-preserving self-evolving optimization paradigm uses paired supervision for the same layout across process conditions to guide intermediate-state recomposition and downstream replay.

Critical strengths
  1. Explicit physical guidance. Flow-Litho uses approximate physical operators, including Abbe imaging and resist response, to guide latent dynamics. Compared with LithoDreamer’s data-derived physical subspaces, this makes the assumptions underlying the correction more explicit and interpretable.

  2. Well-defined correction mechanism. The method combines regional physical gradients through a small, regularized linear system with a unique solution and requires no additional trainable parameters. A norm cap limits the correction magnitude, and the appendix provides detailed derivations.

  3. Carefully designed ablations. The experiments include diagonalizing the Gram matrix, shuffling exchange locations while preserving the number of exchanged pixels, and replacing the correction with a physical loss. These controls help isolate the contributions of individual design choices.

Critical weaknesses
  1. The effects of physical correction are not fully disentangled. The ablations establish the correction module’s contribution, but applying it during both training and inference leaves their respective effects unclear. The GT-based tolerance does not eliminate potential conflicts between the approximate physical prior and accurate predictions. A crossed training and inference ablation, together with correction-before/after comparisons and sensitivity to prior error, would clarify when the correction helps.

  2. Computational efficiency is not established. Although the linear solve is small, constructing it requires decoding, physical-operator evaluations, and regional gradient computations. The absence of additional trainable parameters thus does not establish low overhead. Given the motivation of reducing simulation costs, the paper should report end-to-end inference and planning time, peak memory, and the accuracy–runtime trade-off under matched hardware.

  3. The practical interpretation of process planning needs clarification. The paper should explain how the effective resist threshold maps to a measurable or controllable process quantity and how ADI observations are obtained across these settings. Since each candidate condition also changes the predicted mask, the relationship to tuning a fixed fabricated mask is unclear. Recovery among 36 conditions also provides limited evidence for continuous process optimization or process-window improvement, particularly under measurement noise and near-equivalent conditions.

  4. Metric reporting requires clarification. Reported EPE excludes unmatched gauges without reporting the matching rate, making it difficult to assess the impact of contour failures. A reconstruction baseline for the frozen VAE would also help separate representation error. In addition, the stated definitions imply mIoU ≤ mPA under matched evaluation conditions, yet LMLitho’s ID Resist mIoU is 96.35% in Table 1 while its mPA is 95.23% in Table C.1. This discrepancy requires explanation or correction.

  5. Positioning against related guidance methods and baseline comparability are incomplete. The paper should discuss its relationship to Diffusion Posterior Sampling and flow guidance methods, and compare against an appropriate guidance baseline using the same physical operators. This would better establish the contribution of regional gradient coordination. The LithoDreamer results in Table 1 also match previously published values, so their provenance and the comparability of data splits and evaluation settings should be clarified.

Questions for authors

Please see the weakness section.

3 comments
AuthorsOpenReview ↗

Clarification and Discussion of the Problem

Thanks for your suggestions. We respond to the comments below.

W1: Physical Correction

A: During training, corrected velocities retain GT target-velocity supervision, while correction increments are detached (L230,302-305; Eq.E.5). During inference, corrected velocities govern Euler updates (Eqs.6,13). Training shapes velocity learning; inference changes state evolution.

At https://anonymous.4open.science/r/litho/Physical_Operator.md: 1) Tab.P1 gives a 2×2 training/inference ablation. No correction yields 46.63% Accp / 7.39 nm target ADI EPE; training-only, inference-only, and joint correction yield 58.49%/5.84 nm, 50.94%/6.43 nm, and 78.29%/4.08 nm, respectively. Both stages contribute, and their combination performs best. 2) Tab.P2 compares corrected and uncorrected velocities at matched latent states. One-step GT MSE decreases by 55.28% (Mask), 51.38% (Resist), and 20.26% (ADI), verifying gains in actual GT prediction, not only approximate-operator discrepancy. 3) P and G are separately scaled by ±5%, ±10%, and ±20% (Tab.P3). At ±20%, all four direction-resolved settings (Tab.P4) retain lower ADI EPE and higher Accp than PC-Off. Tab.P5 shows GT MSE reductions in every direction: 27.28-30.29% (Mask), 24.76-26.29% (Resist), and 8.41-9.79% (ADI). Correction retains benefits under tested perturbations, while reduced gains reveal calibration sensitivity.

W2: Cost

A: At https://anonymous.4open.science/r/litho/Overhead.md: Correction adds 97 ms (46.9%) and 4.1 GB (Tab.O1). On the same A100, Flow-Litho vs. LithoDreamer takes 304 vs. 323 ms and 12.3 vs. 14.5 GB, with a 6.81-point higher ADI mIoU (Tab.O2). For 12-rollout GP-BO, component-summed time (rollout plus separately measured search overhead) is 3.79 vs. 4.25 s (10.8% lower), with Accp 78.29% vs. 55.57% (Tab.O3). PC-On vs. PC-Off takes 3.79 vs. 2.68 s; Accp is 78.29% vs. 58.49%, target EPE 4.08 vs. 5.84 nm, and planning peak memory 13.02 vs. 8.79 GB. LithoDreamer planning memory is 15.32 GB. Calibre takes 8.441 vs. 0.304 s (96.4% lower observed runtime;Tab.O4).

W3: Process Planning

A: 1.Image-process pairs are from an actual 55 nm line; ADIs are SEM measurements, not physical-operator outputs (Sec.2.2). T is calibrated from the resist process to parameterize development response; it is an effective threshold, not an independently actuated equipment knob. Flow-Litho helps engineers reduce manual trial-and-error through inverse planning and visualization of fine-grained lithographic state evolution.

2.Flow-Litho supports inference from any process stage to obtain later states. At https://anonymous.4open.science/r/litho/Litho_Process_Planning.md, fixing GT Masks isolates downstream process responses from mask-generation effects. Flow-Litho achieves 98.67% Resist Image mIoU and 87.83% ADI mIoU, with 87.25% Target ADI mIoU and 2.75 nm EPE in ID planning, showing downstream process tuning under a fixed mask (Tabs.L1-L2).

3.Continuous GP-BO improves ID target mIoU/EPE from 80.41%/4.08 nm to 82.46%/3.76 nm. On OOD samples under an unseen joint process condition (T=0.119340), it achieves 76.10%/7.92 nm with 12 rollouts, extending search beyond the 36-point grid (Tab.L3).

4.Removing difference preservation lowers Accp from 78.29% to 65.62% (a 12.67-point drop;Tab.4), supporting its role in exact parameter recovery. Random search yields 93.06% process-response mIoU but only 28.13% Accp (Tab.3), showing that response similarity does not ensure correct parameter identification. With Gaussian white noise added to target ADIs, Accp decreases from 78.29% (clean) to 77.35%, 75.92%, and 75.11% at noise levels of 5, 10, and 15 (Tab.L4).

W4: Metric

A: 1.Our EPE_avg calculation method follows commercial EPE protocols and reports gauge-matching completeness. Flow-Litho achieves 100% at all stages, so no target gauges are excluded from its EPE.

2.Frozen-VAE GT reconstruction achieves 98.95%/99.28%/94.47% mIoU and 0.24/0.15/1.25 nm EPE for Mask/Resist/ADI. Its higher mIoU and lower EPE than full-model predictions indicate that the VAE representation is not the main bottleneck.

3.LMLitho's ID Resist Image metrics contained a copy-paste error. The correct values are 95.23% mIoU and 96.35% mPA, and we've fixed it.

W5: Guidance Method

A: 1.DPS/flow guidance uses likelihood or objective gradients; Flow-Litho uses stage-specific physical discrepancies to correct latent velocities and a Gram matrix to coordinate regional gradients (Eqs.9-13). Under the same split, P/G, 12-rollout budget, and protocol, the adapted baselines have 6.32/9.58 percentage points lower recovery and 3.26/5.81 nm higher target ADI EPE than Flow-Litho (78.29%/4.08 nm), confirming the benefit of regional gradient coordination.

2.We obtained LithoDreamer’s dataset and checkpoint from its authors and ran 10 independent evaluations using identical test sets, rollouts, and metrics, reproducing the published results.

Thank you for the response and for addressing my concerns and it’s unfortunate that discrepancy in the LMLitho metrics is due to the copy and paste error. Additionally for W3 it would be helpful to state explicitly that the fixed-mask experiment searches over 36 observed conditions and that the continuous search predicts a mask for each candidate. Please clarify what these results support about process tuning with an already fabricated mask.

AuthorsOpenReview ↗

Clarification and Discussion of the Problem

Thank you very much for your additional questions. Our responses are provided below.

We model lithography as a three-stage process involving four states: Layout, Mask, Resist Image, and ADI. Flow-Litho supports forward prediction from any intermediate state and target-driven inverse process planning. For an already fabricated mask, its geometry remains fixed while downstream Resist Images and ADIs are predicted under different candidate process conditions. When planning starts from a Layout, each candidate triggers a complete three-stage rollout, re-predicting the Mask and subsequent states. This enables engineers to evaluate and visualize lithographic state evolution before physical trials, reducing trial-and-error. We provide the following additional experiments on fixed-mask planning.

1. Fixed-mask forward prediction. Given an actual GT Mask, Flow-Litho predicts the evolution of subsequent Resist Images and ADIs to evaluate downstream patterning responses under specified exposure conditions. As reported in Table L1 (https://anonymous.4open.science/r/litho/Litho_Process_Planning.md): On 55 nm ID data, Flow-Litho achieves 98.67% and 87.83% mIoU for Resist Image and ADI; on 55 nm OOD data, their EPEs are 0.69 nm and 3.06 nm; on 28 nm cross-node data, Resist Image achieves 90.36% mIoU and 0.74 nm EPE. These results show accurate downstream response prediction under different process conditions with a fixed mask.

2. Fixed-mask discrete inverse planning. Given a fixed GT Mask and a target ADI, GP-BO searches over 36 observed process configurations with 12 candidate rollouts. This enables engineers to identify suitable downstream process conditions for an existing mask without fabricating a new one. As reported in Table L2 (https://anonymous.4open.science/r/litho/Litho_Process_Planning.md): Flow-Litho achieves 82.52% exact parameter recovery, 87.25% target ADI mIoU, and 2.75 nm target ADI EPE. These results show effective downstream process selection for a target ADI while keeping the existing Mask fixed.

3. Fixed-mask continuous inverse planning. We further employ continuous GP-BO, keeping the GT Mask fixed across all candidate evaluations and predicting the corresponding Resist Images and ADIs to optimize downstream process conditions. As shown in Table R1, Continuous GP-BO consistently outperforms Random Search under both ID and OOD settings, achieving better target ADI agreement and process-response consistency. With 24 rollouts, target ADI mIoU further improves to 88.41% and 82.77%, while EPE decreases to 2.39 nm and 3.10 nm. These results demonstrate effective downstream process optimization while preserving the existing mask, enabling process adjustment without mask refabrication.

Table R1. Fixed-Mask Continuous Inverse Process Planning (ID / OOD).

Search MethodRolloutsTarget ADI mIoU (%)Target ADI EPE (nm)Process-Response mIoU (%)Process-Response EPE (nm)Normalized Parameter MAE (%)Source Accuracy (%)Planning Time (s)
Random Search1285.93 / 79.723.26 / 4.6195.45 / 93.921.79 / 2.1312.73 / 16.1787.25 / 84.253.84 / 4.18
Continuous GP-BO1287.25 / 81.692.57 / 3.6897.52 / 95.941.29 / 1.737.53 / 10.6894.05 / 92.364.45 / 4.62
Random Search2486.35 / 80.022.85 / 4.1796.88 / 94.871.68 / 1.904.57 / 6.9596.64 / 94.906.82 / 7.61
Continuous GP-BO2488.41 / 82.772.39 / 3.1098.43 / 96.281.19 / 1.361.23 / 3.2998.69 / 97.167.35 / 7.92

NMAE measures the normalized recovery error of continuous process parameters, while Source Accuracy evaluates categorical illumination-source selection.

4. Full-pipeline continuous inverse planning. When planning starts from a Layout, each candidate undergoes a complete three-stage rollout, independently predicting its corresponding Mask, Resist Image, and ADI under the candidate process conditions. This setting jointly considers mask generation and downstream pattern transfer. As reported in Table L3 (https://anonymous.4open.science/r/litho/Litho_Process_Planning.md), with 12 candidate rollouts, continuous GP-BO achieves 82.46% target ADI mIoU and 3.76 nm EPE on 55 nm ID data, and 76.10% mIoU and 7.92 nm EPE under an unseen joint OOD process condition.

Together, these results establish Flow-Litho's capability for discrete and continuous downstream process planning with an existing mask, without mask refabrication, as well as full-pipeline planning with candidate-specific mask prediction.

We would be pleased to address any further questions or concerns.

Reviewer_zt3fOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a physics-guided world model for multi-stage computational lithography. It uses a shared DiT with flow matching for cross-stage joint learning for faithful multi-stage rollouts. It introduces a physics-guided velocity correction mechanism to improve prediction accuracy and a difference-preserving self-evolving training strategy to reduce error accumulation across stages. Experiments demonstrate strong performance compared with baslines.

Critical strengths

Modeling computational lithography as a multi-stage world model is meaningful, especially for predicting process under different manufacturing conditions.

The design of velocity correction is interesting. It uses physical constraints to guide the flow trajectory and considers gradient conflicts across different image regions.

The method achieves improvements over baselines, particularly under unseen process conditions. The ablation studies are also relatively comprehensive.

Critical weaknesses

Physics-guided sampling and gradient conflict resolution have been studied in previous works. More direct comparisons with alternative physics-guidance strategies would help demonstrate the advantages of the proposed velocity correction.

The difference-preserving recomposition relies on exact pixel equality between ground-truth images under different process conditions. Since some intermediate images contain continuous values, it is unclear how many pixels satisfy this condition.

The proposed method relies on simplified physical operators calibrated offline rather than a complete lithography simulation. If these approximate operators have systematic errors, their gradients may not align with the directions that reduce the true physical errors. How robust is the proposed velocity correction to such inaccuracies? Could the authors provide a sensitivity analysis for this?

Questions for authors

What is the additional inference time introduced by the physical correction module? How does the overall computational cost compare with LithoDreamer and conventional lithography simulators?

How sensitive is the method to inaccuracies in the approximate physical operators?

1 comment
AuthorsOpenReview ↗

Clarification and Discussion of the Problem

Thank you for your suggestions. We respond to your comments below.

W1: Velocity Correction.

A: We compare Flow-Litho with Vanilla Gradient Guidance, PCGrad, and ConFIG on the same 55 nm ID split and evaluation protocol. They differ only in gradient combination: direct summation, pairwise conflict projection, conflict-free integration, and the coupled Gram-matrix solve (Eq.12).

MethodADI mIoUADI EPEExact Parameter Recovery (Accp) (%)
Vanilla Gradient Guidance70.1812.8262.63
PCGrad-based Correction77.436.1768.89
ConFIG-based Correction79.154.8270.05
Ours85.083.6778.29

On this 55 nm ID evaluation, Flow-Litho achieves the best results on all three metrics, improving ADI reconstruction and process-parameter identification through regional gradient coordination (Sec.3.2, Eqs.(9)-(13)).

Compared with physical-loss supervision, Flow-Litho reduces ADI forward EPE from 4.47 to 3.67 nm and raises Accp from 58.35% to 78.29% (Tab.C.4). Removing cross-region coupling lowers Accp to 54.74% (-23.55 percentage points; Tab.4), supporting the role of regional coordination in identifying process-dependent responses.

W2: Pixel Equality Details

A: Exact equality is assessed by direct pixel-wise comparison, without thresholding or approximation, of spatially aligned stored GT labels at the Mask and Resist Image stages. Recomposition exchanges predicted states only at matching pixels; mismatches retain their branch predictions (L 317-323, 1421-1424).

Training-set statistics show that exact matches account for 94.37%/96.58% of common valid pixels and 90.44%/89.66% of Eq.(9) boundary bands in the Mask/Resist Image stages. The 0.5 threshold constructs the boundary mask and is not used for the equality test in Eq.(14).

Removing recomposition lowers Accp by 11.83 percentage points (Tab.C.5); permuting the exchange mask while preserving per-quadrant counts lowers it by 12.67 percentage points (Tab.4). Results show that state exchange and GT-label-consistent exchange locations contribute to accurate process-parameter identification.

W3, Q2: Physical Operators.

A: Approximate operators serve as physical references rather than exact lithography simulators. Flow-Litho limits their influence through GT-calibrated regional tolerances based on the 90th percentile of GT-pair discrepancies, a correction-norm bound of 20% of the base-velocity norm, and GT target-velocity supervision with detached correction increments (Eqs.11-13; L287-305; Eq.E.5).

For sensitivity evaluation (https://anonymous.4open.science/r/litho/Physical_Operator.md), the checkpoint and original regional tolerances are fixed. P and G are perturbed separately by jointly scaling all coefficients of the selected operator by 1+δ or 1−δ, where δ∈{0.05, 0.10, 0.20}; the other operator remains unchanged.

SettingID Mask EPEID Resist Image EPEID ADI EPEAccpGT-Dir. (%)
PC-Off (Training ON/Inference OFF)1.381.154.7858.49--
Original P and G0.420.373.6778.2998.42
P, ±5/10/20%0.75/0.93/1.240.68/0.87/1.113.79/4.08/4.4273.52/70.49/65.8591.53/81.02/76.47
G, ±5/10/20%0.420.373.73/3.93/4.3674.06/71.34/67.1793.29/86.54/76.63

(1) At ±20%, direction-resolved results (Tab.P4) show ADI EPE/Accp of 4.45/66.88% and 4.39/64.82% for P +20%/−20%, and 4.35/67.18% and 4.37/67.16% for G +20%/−20%. All four settings retain lower ADI EPE and higher Accp than PC-Off (4.78 nm and 58.49%). (2) One-step GT image errors (Tab.P5) provide a direct check of correction effects. Under P +20%/−20%, Mask MSE decreases from 27.93 to 19.47/20.31 and Resist Image MSE from 24.27 to 17.89/18.26. Under G +20%/−20%, ADI MSE decreases from 30.45 to 27.47/27.89 (×1e-3). These correspond to relative reductions of 30.29% and 27.28% for Mask, 26.29% and 24.76% for Resist Image, and 9.79% and 8.41% for ADI. (3) GT-Dir. decreases from 98.42% with the original operators to 76.47% for P and 76.63% for G at ±20% (Tab.P3). Thus, the correction retains measurable stage-wise GT-error reduction and process-planning benefits under the tested perturbations.

Q1: Inference Cost.

A: At https://anonymous.4open.science/r/litho/Overhead.md, we quantify physical correction overheads: PC-On and PC-Off use the same checkpoint, inputs, and nine Euler updates; neural-model timings use one A100 80GB, batch size 1, and 512×512 inputs. Correction adds 97 ms (46.9%) mean latency and 4.1 GB peak allocated memory relative to inference with correction disabled (Tab.O1). Despite this overhead, the complete Flow-Litho achieves 5.9% lower latency and 2.2 GB lower peak allocated memory than LithoDreamer, while improving ADI mIoU by 6.81 percentage points (Tab.O2). Compared with the commercial simulator, Flow-Litho completes the three-stage workflow in 0.304 s vs. 8.441 s, reducing runtime by 96.4% (Tab.O4).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.