PaperlayerView exact revision on arXiv

Source-linked AI summary

VPEvolve: A Self-Evolving Virtual Process Engineer for Computational Lithography

Tianyi Li, Wenxuan Dong, Donger Luo, Nan Wang, Yanpeng Chen, Jiaqi Liu, Xinyun Zhang, Hao Geng

arXiv:2609.32473v1cs.AIcs.SE

TL;DR

OPC recipes grow with hotspot-specific rules that can interact, while lessons from commercial-tool trials remain scattered. VPEvolve combines a frozen-model engineering harness with a Skill Bank that turns measured responses into judgments for later trials. On a FreePDK45-derived benchmark, it lowers mean per-case maximum EPE on Poly and Metal1 while every endpoint meets predefined quality constraints.

  • Problem

    OPC recipes accumulate interacting local rules, while the reasons and outcomes of commercial-tool trials remain scattered and difficult to reuse.

    Source: §1 · p. 2

  • Method

    VPEvolve equips a frozen LLM with process guidance, layout analysis, recipe editing, and commercial-tool evaluation, while reflection and curation convert trial outcomes into evidence-linked judgments for subsequent actions.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    18.294 to 5.361 nm on Poly and 22.052 to 15.692 nm on Metal1 for mean window-maximum EPE; all cases achieve a feasible improvement.

    Source: §4.2 · pp. 6–7

  • Takeaways & Limitations

    The released FreePDK45 layouts and lithography simulator support open simulation studies.

    Source: §1 · p. 2

Abstract

from arXiv · show

Optical proximity correction (OPC) recipes grow as engineers add local rules to repair newly discovered lithography hotspots. Each correction can interact with existing rules, while lessons from commercial-tool trials remain scattered across code and logs. \system combines a Virtual Process Engineer (VPE) harness with a Skill Bank of measured engineering experience. The harness equips a frozen language model with process manuals, layout analysis, recipe editing, and commercial-tool evaluation. The actor proposes changes to the global parameters, local targeted rules, or diagnostic trials. After each evaluation, an LLM reflector and curator turn the measured response into evidence-linked judgments. The actor retrieves them before its next trial. Feasible improvements update the retained recipe; every measured trial informs the Skill Bank that guides the next edit. The model weights remain fixed. On a FreePDK45-derived benchmark with ten commercial-tool evaluations per case, \system reduces the mean per-case maximum edge placement error from 18.294 to 5.361 nm on Poly and from 22.052 to 15.692 nm on Metal1. Every final recipe satisfies the predefined quality constraints and improves the maximum error by at least 0.1 nm.

1 INTRODUCTION

Growing local OPC rules can interact, while trial lessons remain scattered and difficult to reuse. VPEvolve combines a Virtual Process Engineer harness with a Skill Bank that turns measured outcomes into reusable judgments.

1 INTRODUCTION

  • Recipes include thousands of local-rule lines that target recurring patterns, and a fix can interact with existing rules or create another violation.The link between what changed, why, and what happened remains scattered across engineers, code, and logs.Source: §1 · p. 2
  • Parameter tuning searches settings chosen in advance, whereas new hotspots can require local rules outside that search space and can change how earlier trial lessons apply.A rule may repair its target while another hotspot becomes the limiting error.Source: §1 · p. 2
  • VPEvolve equips a frozen LLM with process manuals, layout analysis, recipe editing, and commercial-tool feedback, while reflection and curation turn trial outcomes into evidence-linked judgments for later edits.The actor can propose global edits, targeted local corrections, or diagnostic trials.Source: §1 · p. 2
  • 18.294 to 5.361 nm on Poly and 22.052 to 15.692 nm on Metal1 for mean per-case maximum EPE; every endpoint meets quality constraints and improves by at least 0.1 nm.The benchmark allows ten commercial-tool evaluations per case.Source: §1 · p. 2
  • The paper frames recipe development as sequential search over global settings and growing local rules, and reports lower maximum EPE than all seven baselines on both layers.The FreePDK45-derived layouts and lithography simulator are released for open simulation studies.Source: §1 · p. 2

2 PRELIMINARIES

The paper evaluates OPC recipes by simulating printed contours across focus and exposure conditions, then comparing contours with target edges and assessing process variation. It situates this task among prior parameter-search, lithography-learning, and agent-memory approaches.

2 PRELIMINARIES

  • A recipe guides SRAF insertion and OPC to produce a mask, whose printed contours are simulated across focus and exposure conditions; EPE compares contours with target edges and PVB measures their variation.These measurements guide subsequent recipe edits.Source: §2 · pp. 2–3, Figure 2 · p. 3
  • Reported Max and Mean EPE are maximum and mean absolute errors at fixed gauges under nominal focus and dose, while Top-10 EPE averages ten highest-error spatially separated hotspots.PVB is normalized by target-edge length; MRC and SRAF printability checks assess manufacturability and unintended printing.Source: §2 · p. 3
  • Prior Bayesian optimization searches a parameter space specified in advance, while lithography-learning work includes mask corrections, SRAFs, and learned placement of EPE measurements and edge fragmentation.Recipes also grow through hotspot-driven local rules, which can change how earlier trial outcomes inform later decisions.Source: §2 · p. 3
  • Prior agents coordinate tool use and experiments, while reflection, memory, skills, and playbooks carry information across attempts; VPEvolve evolves both recipes and engineering judgments from measured feedback.The agent harness stays fixed.Source: §2 · pp. 3–4

3 METHOD

VPEvolve treats recipe development as sequential decisions under a fixed evaluation process, retaining feasible recipe improvements while preserving trial evidence. Its harness supports global, local, and diagnostic actions, and its Skill Bank converts measured responses into scoped judgments and suggested experiments.

3.1 RECIPE DEVELOPMENT AS A SEQUENTIAL DECISION PROBLEM

  • The search aims to reduce maximum EPE while maintaining acceptable average EPE, process variation, and manufacturability, with the process model and evaluation settings fixed.At each step, a candidate replaces the retained recipe only if it satisfies quality constraints and improves the objective; otherwise, the current recipe remains.Source: §3.1 · p. 4
  • Every evaluation records the change, conditions, and printed response; the actor retrieves relevant judgments, while a reflector interprets the outcome and a curator proposes bank revisions.Trial records remain available even when a candidate recipe is rejected, allowing evidence to inform later decisions independently of recipe retention.Source: §3.1 · p. 4

3.2 VIRTUAL PROCESS ENGINEER HARNESS

  • The harness connects a frozen LLM actor to process guidance, layout analysis, recipe editing, and commercial-tool evaluation; the actor chooses global edits, local corrections, or diagnostic trials.The figure describes feasible improvements updating the retained recipe while reflection and curation revise experience for subsequent trials.Source: §3.2 · p. 4, Figure 3 · p. 5
  • For local actions, layout tools associate hotspots with design edges and nearby geometry, and a deterministic compiler binds edits to their parent recipe and checks commands, geometry, and displacement limits.Commercial-tool measurements include hotspot and global metrics plus process-variation and manufacturability checks; retention follows fixed constraints and a lexicographic quality ordering.Source: §3.2 · p. 5

3.3 SELF-EVOLVING SKILL BANK

  • Domain Response Reflection compares intended and measured effects at the target hotspot, new worst hotspot, and scoring region; Scoped Experience Revision links judgments to applicability conditions, supporting trials, counterexamples, and unresolved questions.Experiment Guidance proposes up to two next experiments, each with expected effects and a falsifying observation; the actor chooses the next action.Source: §3.3 · p. 5
  • The actor uses retained records and distilled findings in later decisions without changing model weights or harness code.The table caption describes layer-wise EPE and PVB case means, with benchmark-wide success and call counts.Source: §3.3 · p. 6, Table 1 · p. 6

4 EXPERIMENTS

Across ten commercial-tool evaluations per case, VPEvolve improves recipe quality on both benchmark layers. Ablations and extended trials examine experience evolution, local correction, sustained refinement, and performance across language models.

4.1 EXPERIMENTAL SETUP

  • The benchmark covers 20 FreePDK45-derived windows, with ten trials per case and constraints on mean EPE, PVB, MRC, and SRAF printability.Baselines include agent and memory methods plus Bayesian optimization; implementations use fixed model weights.Source: §4.1 · p. 6

4.2 MAIN RESULTS

  • VPEvolve lowers mean window-maximum EPE from 18.294 to 5.361 nm on Poly and from 22.052 to 15.692 nm on Metal1, with every case achieving a feasible improvement.Mean EPE also falls below every baseline, to 0.645 nm on Poly and 0.690 nm on Metal1.Source: §4.2 · pp. 6–7

4.3 MECHANISM ABLATION

  • Frozen Experience reaches 8.082 nm on Poly and 16.121 nm on Metal1, while generic reflection reaches 6.213 and 16.280 nm, respectively.The Frozen variant omits earlier trial history and online experience updates; generic reflection retains spatial evidence but removes explicit guidance for interpreting hotspot responses and overall recipe quality.Source: §4.3 · p. 7
  • Removing scoped revision raises maximum EPE from 5.361 to 5.475 nm on Poly and from 15.692 to 16.334 nm on Metal1; removing experiment guidance yields 5.870 and 16.154 nm.On Poly, removing revision or guidance lowers mean EPE from 0.645 nm to 0.632 or 0.624 nm, showing worst-hotspot and average accuracy can diverge.Source: §4.3 · pp. 7–8

4.4.1 VALUE OF EVOLVED EXPERIENCE

  • With recipe and history fixed, VPEvolve experience yields mean additional maximum-EPE reductions of 0.681 nm on Poly and 2.494 nm on Metal1, versus 0.453 and 2.104 nm for ACE experience.Both experience banks remain fixed during five continuation trials from the shared fifth-trial recipe and history.Source: §4.4.1 · p. 8, Figure 5 · p. 8

4.4.2 LOCAL CORRECTION AFTER GLOBAL EXPLORATION

  • After four paired trials, allowing local actions lowers mean maximum EPE from 18.583 to 18.012 nm on Metal1 and from 5.238 to 5.185 nm on Poly.Mean EPE and PVB are higher with local actions, though both branches remain feasible and the value depends on remaining hotspots and available global edits.Source: §4.4.2 · p. 8, Figure 5 · p. 8

4.4.3 LONG-HORIZON OPTIMIZATION

  • In 100-trial Poly runs, Full reaches 4.477 nm maximum EPE versus 5.267 nm for Frozen Experience; on Metal1, Full reaches 10.912 nm while Frozen terminates at trial 57.At that shared Metal1 horizon, Full already has lower maximum EPE: 12.470 versus 13.348 nm.Source: Figure 6 · p. 8, §4.4.3 · p. 9

4.4.4 GENERALIZATION ACROSS LANGUAGE MODELS

  • All six tested language models lower mean and maximum EPE from R0, with model rankings varying by layer and metric.Mean maximum EPE ranges from 7.661 to 10.258 nm on Poly and 16.696 to 17.398 nm on Metal1.Source: §4.4.4 · p. 9, Figure 7 · p. 9

5 CONCLUSION

VPEVOLVE treats each lithography trial as both a recipe-improvement opportunity and evidence for the next engineering decision. Its frozen-LLM harness and Skill Bank support this loop, which the authors evaluate with commercial-tool experiments and an open simulation testbed.

5 CONCLUSION

  • VPEVOLVE makes each lithography trial part of a loop that improves a recipe and leaves evidence for the next engineering decision.The model weights remain fixed while the Skill Bank turns trial feedback into scoped, evidence-linked judgments.Source: §5 · p. 9
  • A frozen LLM harness combines process guidance, layout analysis, recipe editing, and commercial-tool evaluation with reflection and curation of trial experience.The released FreePDK45 layouts and simulator support open simulation studies.Source: §5 · p. 9

A LAYOUTS AND DATASET CONSTRUCTION

The benchmark uses cropped FreePDK45 full-chip layouts to evaluate Poly and Metal1 windows across five density strata. Each window has a fixed scoring core and context halo, and the layouts and metadata are publicly released.

A LAYOUTS AND DATASET CONSTRUCTION

  • Each layer has ten FreePDK45-derived windows, with two windows from each of five density strata spanning very sparse to very dense layouts.Poly is layer 9/0 and Metal1 is layer 11/0; each window is one evaluation unit in the main comparison.Source: A LAYOUTS AND DATASET CONSTRUCTION · p. 13
  • Each window retains whole polygons intersecting a 2 µm context halo, but scores only its fixed 10 × 10 µm core.The sampling manifest records the source-layout hash, layer, density group, and core and context coordinates.Source: A LAYOUTS AND DATASET CONSTRUCTION · p. 13
  • The benchmark layouts and their metadata are included in the public release.The local-correction study uses one case per density stratum in each layer.Source: A LAYOUTS AND DATASET CONSTRUCTION · p. 13

B EXPERIMENTAL PLATFORM AND EVALUATION PROTOCOL

The evaluation fixes the process model, rule decks, scoring code, allowed edits, and evaluation budget while the actor selects trials and measured responses guide recipe and experience updates. The protocol defines feasibility and selection criteria, then compares the workflow with agent and numerical-search baselines.

B EXPERIMENTAL PLATFORM AND EVALUATION PROTOCOL

  • Experiments use licensed Calibre with two threads per job and fixed process, rule-deck, scoring, edit, and budget settings.Candidate validation checks the parent recipe, command syntax, and allowed edits, with added pattern-selection and spatial-motion checks for local corrections.Source: B EXPERIMENTAL PLATFORM AND EVALUATION… · p. 13
  • A feasible recipe has zero MRC and SRAF-printing violations, with mean EPE and PVB within 102% of R0.Selection minimizes maximum EPE, breaking ties by Top-10 EPE, Mean EPE, and PVB, in that order.Source: B EXPERIMENTAL PLATFORM AND EVALUATION… · p. 13
  • Success requires a completed run with a feasible maximum-EPE reduction of at least 0.1 nm; quality means weight assigned windows equally and use each run’s last retained feasible recipe.Missing measurements are not imputed.Source: B EXPERIMENTAL PLATFORM AND EVALUATION… · p. 13
  • Trial budgets include evaluated candidates that are not retained and execution or measurement failures, while rejected pre-evaluation proposals consume no OPC evaluation.Initialization and endpoint repeats are separate from search; Table A1 lists configurations and per-run budgets, excluding initialization and final reevaluation.Source: B EXPERIMENTAL PLATFORM AND EVALUATION… · p. 13, Table A1 · p. 14

C BASELINE DETAILS

  • The baselines cover direct action selection, interaction history, accumulated experience, and numerical search under shared initial recipes, budgets, and quality constraints.Agent baselines use the same recipe-editing actions.Source: C BASELINE DETAILS · p. 13
  • Raw Actor uses measurements without an explicit experience bank, while ReAct uses interaction history and tool feedback to revise its working hypothesis.Neither baseline has a separate experience bank.Source: C BASELINE DETAILS · p. 13
  • Textual Memory summarizes prior attempts chronologically, whereas ReasoningBank retrieves reusable lessons using lexical overlap with the case and hotspot feedback.ReasoningBank lessons include applicability conditions and potential pitfalls.Source: C BASELINE DETAILS · p. 14
  • ACE incrementally curates a playbook, Static Harness keeps engineering guidance fixed, and Bayesian Optimization searches model-based parameters with Gaussian-process surrogates.Bayesian Optimization selects candidates using constrained expected improvement and updates its surrogates after each evaluation.Source: C BASELINE DETAILS · p. 14

D IMPLEMENTATION AND EXPERIMENTAL SETTINGS

VPEVOLVE uses a frozen model for its actor, reflector, and curator, with shared initialization and model settings for the main comparison, ablations, and paired continuations. Its optimization loop retrieves experience, evaluates trials, retains feasible improvements, and revises the Skill Bank from measured responses.

D IMPLEMENTATION AND EXPERIMENTAL SETTINGS

  • Unless otherwise specified, all LLM roles use Qwen3.6-27B locally at temperature zero, with a 131,072-token context and 4,096 output tokens.VPEvolve and its mechanism ablations start from R0 with an empty Skill Bank and shared engineering guidance.Source: D IMPLEMENTATION AND EXPERIMENTAL… · p. 14
  • In the cross-model study, all three roles use one evaluated model through its API, with a 65,536-token context and provider-specific generation settings.Long-horizon runs use DeepSeek-V4.1-Flash for all three roles through its API.Source: D IMPLEMENTATION AND EXPERIMENTAL… · p. 14
  • Each optimization step retrieves relevant experience, chooses a global, local, or diagnostic trial, evaluates it, and retains the candidate only if it is a feasible improvement.The loop then reflects on the trial, curates the Skill Bank, and re-evaluates the retained endpoint after the budget.Source: D IMPLEMENTATION AND EXPERIMENTAL… · p. 15
  • Reflection compares predicted and measured responses at fixed hotspots, new worst locations, and aggregate metrics to distinguish ineffective edits from repaired targets followed by new limiting hotspots.Curation can add, update, split, or merge judgments linked to supporting trials and counterexamples, and may suggest experiments with predicted effects and falsifying observations.Source: D IMPLEMENTATION AND EXPERIMENTAL… · p. 15
  • Frozen Experience fixes the initial bank and guidance, while ablations remove reflection guidance, evidence-review mechanisms, or strategy updates and proposed experiments.All experience-update conditions share actor guidance, available actions, physical selection criteria, and factual citation checks.Source: D IMPLEMENTATION AND EXPERIMENTAL… · p. 15

E PAIRED CONTINUATION STUDIES

The paired continuation studies compare the value of VPEvolve-curated experience and allowing local actions, using matched starting conditions. Both alternatives show positive mean advantages on both layers, with local actions trading off higher mean EPE and PVB within feasibility limits.

E PAIRED CONTINUATION STUDIES

  • The experience study resumes all 20 cases from Full’s fifth-trial recipe, feedback, and history, then compares fixed banks over five further trials.ACE curates the five preceding observations; VPEvolve supplies recorded experience, with shared model settings, engineering guidance, and trial history.Source: E PAIRED CONTINUATION STUDIES · p. 15
  • The local-action study compares global-only edits with global and local edits from matched states over four trials, with selected mask-fragment displacements checked before execution.It selects five Poly and five Metal1 cases before observing outcomes while keeping target geometry and evaluation gauges fixed.Source: E PAIRED CONTINUATION STUDIES · p. 15
  • Endpoint quality reports EPE in nm and PVB in 10−3 µm; experience conditions cover the benchmark, while each local-action condition uses five cases per layer.Source: Table A3 · p. 16
  • VPEvolve experience and local actions have positive mean maximum-EPE advantages on both layers; local actions also increase mean EPE and PVB within feasibility limits.Source: E PAIRED CONTINUATION STUDIES · p. 15, Table A4 · p. 16

F EVALUATION AND MODEL-CALL COST

The main comparison uses the same commercial-tool evaluation budget across its assigned windows, while VPEvolve makes fewer model requests per window than ACE but more than ReAct.

F EVALUATION AND MODEL-CALL COST

G LONG-HORIZON EVALUATION

The long-horizon evaluation compares Full and Frozen variants from R0 with empty Skill Banks. Full completes both layer runs, while Frozen’s Metal1 run ends after repeated requests for evidence excluded by its history restriction.

G LONG-HORIZON EVALUATION

  • Both variants start from R0 with empty Skill Banks and shared guidance on one case per layer, using the stated API settings and evaluation accounting.Source: G LONG-HORIZON EVALUATION · p. 16
  • Across both layers, Full and Frozen complete 200 and 157 trials (357 total), excluding initialization and endpoint repeats.Source: G LONG-HORIZON EVALUATION · p. 16
  • Both Poly runs and Metal1 Full complete 100 trials, but Metal1 Frozen terminates at trial 57 after repeated restricted-evidence requests and recovery.Poly Full repeats evidence requests after trial 97 but completes after recovery; recovery extends model-call allowances, with actual costs retained.Source: G LONG-HORIZON EVALUATION · p. 16
  • The Metal1 quality comparison measures Frozen at trial 57 and records further gains as Full continues.Source: G LONG-HORIZON EVALUATION · p. 16

Related papers

Loading 2609.32473v1…