pith. sign in

arxiv: 2606.00923 · v1 · pith:573RUNCQnew · submitted 2026-05-30 · 💻 cs.CE

Graph Attention-Based Virtual Metrology for Film Deposition Processes in Semiconductor Manufacturing

Tao Han , Suk Ki Lee , Hyunwoong Ko This is my paper

Pith reviewed 2026-06-28 17:35 UTC · model grok-4.3

classification 💻 cs.CE
keywords virtual metrologygraph attention networkssemiconductor manufacturingfilm depositionprocess monitoringpredictive modelinginterpretability
0
0 comments X

The pith

A graph attention model predicts semiconductor film thickness from sensor data more accurately than baselines while showing interpretable parameter relationships.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a graph attention framework for virtual metrology in film deposition that turns each step-parameter pair into a graph node and uses convolutional encoders to pull temporal features from equipment traces. A parameter-to-layer attention step then lets each film layer gather relevant directional information from the process variables. The goal is to forecast wafer film thickness without waiting for physical measurements, which are slow and expensive at high volume. On industrial production data the model outperforms standard baselines, and the learned attention weights line up with known physical drivers and timing across deposition stages. If these results hold, manufacturers could monitor and adjust processes with less reliance on sampled physical checks.

Core claim

The framework represents process steps and parameters as nodes, extracts temporal embeddings via convolutional encoders, and applies a parameter-to-layer graph attention mechanism so each film layer aggregates the most relevant process signals; on industrial deposition data this yields higher prediction accuracy for film thickness than baseline models while the attention weights recover dominant factors and temporal patterns that match physical process behavior.

What carries the argument

The parameter-to-layer graph attention mechanism, which builds a directed graph over step-parameter nodes and lets each film layer selectively aggregate temporal features from relevant upstream parameters.

If this is right

  • The model achieves better predictive performance than baseline models when tested on the collected industrial deposition data.
  • Attention weights identify dominant process factors and temporal dependencies that align with physical deposition behavior.
  • The framework supplies both thickness predictions and process insights that can support monitoring and optimization decisions.
  • Temporal feature extraction combined with structured graph dependencies reduces the impact of heterogeneous sensor variables on prediction quality.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If attention weights prove stable across tools, the same graph structure could be reused to flag which sensor channels to prioritize during equipment maintenance.
  • Extending the node definition to include upstream etch or clean steps might allow end-to-end virtual metrology across multiple process modules without retraining from scratch.
  • Real-time sensor streams fed through the same encoders could support closed-loop thickness control rather than post-process prediction only.

Load-bearing premise

The industrial production-wafer dataset is representative and unbiased enough that performance gains and attention interpretations will hold on new tools or process variations.

What would settle it

Running the same model on deposition data from a different tool or with deliberate process shifts and finding that prediction error does not beat baselines or that attention weights no longer match documented physical dependencies.

Figures

Figures reproduced from arXiv: 2606.00923 by Hyunwoong Ko, Suk Ki Lee, Tao Han.

Figure 1
Figure 1. Figure 1: FIGURE 1: OVERALL FRAMEWORK OF THE PROPOSED METHODOLOGY [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIGURE 2: TRAINING AND VALIDATION LOSS CURVES OVER 100 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIGURE 4: FULL HEATMAP OF ATTENTION WEIGHTS AVERAGED [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIGURE 5: HEATMAP OF THE TOP-K ATTENTION WEIGHTS [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIGURE 6: BAR CHARTS OF THE TOP- [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Artificial intelligence-driven semiconductor manufacturing increasingly operates at nanometer and angstrom scales, where precise process control depends on accurate and timely metrology. However, physical metrology is limited by measurement latency, cost, and sampling constraints, restricting its scalability in high-volume production. Virtual metrology (VM) has emerged as an effective alternative by predicting wafer-level characteristics from equipment sensor data. Despite recent advances, many existing VM models remain correlation-driven and lack the ability to capture structured dependencies among heterogeneous process variables, while providing limited interpretability. This study presents a graph attention-based VM framework for film deposition processes that integrates temporal feature learning with structured parameter-layer dependency modeling. The proposed approach represents each step-parameter pair as a node and extracts temporal embeddings from high-frequency equipment traces using convolutional feature encoders. A parameter-to-layer graph attention mechanism is employed to model directional dependencies, enabling each film layer to aggregate relevant process information. The framework is evaluated using industrial deposition data collected from production wafers, where the model predicts film thickness from multivariate sensor signals. Experimental results demonstrate improved predictive performance compared to baseline models. In addition, analysis of the learned attention weights reveals interpretable parameter-layer relationships consistent with physical process behavior, capturing dominant process factors and temporal dependencies across deposition stages. These results indicate that the proposed framework enhances prediction accuracy and provides meaningful insight into process dynamics, supporting effective monitoring and optimization in semiconductor manufacturing.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents a graph attention-based virtual metrology framework for film deposition processes in semiconductor manufacturing. It models each step-parameter pair as a node, extracts temporal embeddings from high-frequency equipment traces via convolutional encoders, and employs a parameter-to-layer graph attention mechanism to model directional dependencies so that each film layer aggregates relevant process information. Evaluated on industrial deposition data from production wafers, the model predicts film thickness; the manuscript claims improved predictive performance over baselines plus interpretable attention weights that capture dominant process factors and temporal dependencies consistent with physical behavior.

Significance. If the performance gains and physical consistency of the attention weights are substantiated with quantitative evidence, the work could advance virtual metrology by combining structured dependency modeling with interpretability, aiding process monitoring and optimization where physical metrology is costly or slow.

major comments (2)
  1. [Experimental Results] Experimental Results section: the claim of improved predictive performance supplies no quantitative metrics, baseline descriptions, error bars, dataset size, or validation protocol, so the central empirical claim cannot be evaluated.
  2. [Attention Weight Analysis] Attention weight analysis: the claim that learned attention weights reveal 'interpretable parameter-layer relationships consistent with physical process behavior' rests on post-hoc inspection alone; no ablation removing known physical drivers, no comparison to a physics-based reference model, and no out-of-distribution test on a different tool or recipe are described, leaving open the possibility that patterns reflect training-distribution correlations rather than causal dependencies.
minor comments (1)
  1. [Abstract] Abstract: the statement 'experimental results demonstrate improved predictive performance' should include at least a high-level reference to the specific metrics or tables that support it.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to strengthen the presentation of results and analysis.

read point-by-point responses

  1. Referee: [Experimental Results] Experimental Results section: the claim of improved predictive performance supplies no quantitative metrics, baseline descriptions, error bars, dataset size, or validation protocol, so the central empirical claim cannot be evaluated.

    Authors: We agree that the Experimental Results section in the submitted manuscript omits these essential details, which prevents evaluation of the performance claims. In the revised manuscript we will expand the section to report quantitative metrics including MAE and RMSE with error bars across runs, describe the baseline models, state the dataset size (number of production wafers and sensor traces), and specify the validation protocol such as the train/test split or cross-validation procedure. revision: yes

  2. Referee: [Attention Weight Analysis] Attention weight analysis: the claim that learned attention weights reveal 'interpretable parameter-layer relationships consistent with physical process behavior' rests on post-hoc inspection alone; no ablation removing known physical drivers, no comparison to a physics-based reference model, and no out-of-distribution test on a different tool or recipe are described, leaving open the possibility that patterns reflect training-distribution correlations rather than causal dependencies.

    Authors: The current analysis is indeed limited to post-hoc inspection of attention weights for consistency with known process physics. We will add an ablation study that removes parameters identified as dominant by the attention mechanism and quantifies the resulting change in prediction error and attention patterns. Direct comparison to a physics-based reference model is not straightforward for this deposition process, but we will provide a more detailed mapping to process engineering knowledge. Out-of-distribution evaluation on a different tool or recipe is not possible with the available industrial dataset. revision: partial

Circularity Check

0 steps flagged

No significant circularity; model trained on external data with independent evaluation

full rationale

The paper describes an end-to-end graph attention neural network trained on industrial deposition sensor traces to predict film thickness. Reported performance metrics are standard held-out test errors from this external dataset; no equations define the predictions in terms of the fitted parameters themselves, and no self-citation chain or ansatz is invoked to justify the central results. Attention-weight analysis is presented as post-hoc inspection rather than a load-bearing derivation. The framework therefore remains self-contained against external benchmarks with no reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The central claim rests on standard supervised learning assumptions plus the domain premise that sensor traces contain sufficient structured information to reconstruct layer thicknesses and that a graph can faithfully encode the relevant physical dependencies.

free parameters (1)
  • model hyperparameters
    Number of attention heads, embedding dimensions, learning rate, and graph construction thresholds are chosen or fitted during training and affect the reported performance.
axioms (2)
  • domain assumption Sensor data from equipment traces contains the information needed to predict film thickness at each layer
    Invoked when the framework is presented as an effective alternative to physical metrology.
  • domain assumption Directional dependencies between process parameters and film layers can be represented as a directed graph
    Central to the parameter-to-layer graph attention mechanism described in the abstract.
invented entities (1)
  • parameter-to-layer graph attention mechanism no independent evidence
    purpose: To allow each film layer to selectively aggregate information from relevant process parameters via learned attention weights
    New modeling construct introduced to capture structured dependencies that prior correlation-driven VM models lack.

pith-pipeline@v0.9.1-grok · 5776 in / 1346 out tokens · 25445 ms · 2026-06-28T17:35:28.615484+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 2 canonical work pages

  1. [1]

    Silicon minds: The rise of AI- powered chips

    Talati, Dhruvitkumar. “Silicon minds: The rise of AI- powered chips.” (2021)

    2021

  2. [2]

    Development trendsandprospectsofsemiconductordevicesandtechnol- ogy

    Qiu, Zekun, Shen, Xianao and Zhao, Zirui. “Development trendsandprospectsofsemiconductordevicesandtechnol- ogy.”DevelopmentVol. 81 (2024)

    2024

  3. [3]

    Nanoscale technologies: design challenges and advance- ments

    Srivastava, Sumit, Jaiswal, Abhinav and Khan, Arman. “Nanoscale technologies: design challenges and advance- ments.”Semiconductor Nanoscale Devices: Materials and Design Challenges. Bentham Science Publishers (2025): pp. 1–26

    2025

  4. [4]

    Generative Model PredictiveControlinManufacturingProcesses: AReview

    Lee,SukKi,Stone,RonnieFP,Gao,Max,Zhang,Wenlong, Sha, Zhenghui and Ko, Hyunwoong. “Generative Model PredictiveControlinManufacturingProcesses: AReview.” arXiv preprint arXiv:2511.17865(2025)

    arXiv 2025

  5. [5]

    A review of thin-film growth, properties, applications, and future prospects

    Sakthinathan, Subramanian, Meenakshi, Ganesh Abi- naya, Vinothini, Sivaramakrishnan, Yu, Chung-Lun, Chen, Ching-Lung, Chiu, Te-Wei and Vittayakorn, Naratip. “A review of thin-film growth, properties, applications, and future prospects.”ProcessesVol. 13 No. 2 (2025): p. 587

    2025

  6. [6]

    Cambridge Univer- sity Press (2023)

    Plummer, James D and Griffin, Peter B.Integrated Circuit Fabrication: Science and Technology. Cambridge Univer- sity Press (2023)

    2023

  7. [7]

    Metrology for the next generation of semiconductor 8 FIGURE 6: BAR CHARTS OF THE TOP-k(k=10) PARAMETERS FOR SELECTED LAYERS devices

    Orji, Ndubuisi G, Badaroglu, Mustafa, Barnes, Bryan M, Beitia, Carlos, Bunday, Benjamin D, Celano, Umberto, Kline, Regis J, Neisser, Mark, Obeng, Yaw and Vladar, AE. “Metrology for the next generation of semiconductor 8 FIGURE 6: BAR CHARTS OF THE TOP-k(k=10) PARAMETERS FOR SELECTED LAYERS devices.”Nature electronicsVol. 1 No. 10 (2018): pp. 532–547

    2018

  8. [8]

    7/5nm logic manufacturing capabilities and requirements ofmetrology

    Bunday,Benjamin,Bello,AF,Solecky,EricandVaid,Alok. “7/5nm logic manufacturing capabilities and requirements ofmetrology.”Metrology,Inspection,andProcessControl forMicrolithographyXXXII,Vol.10585: pp.81–124.2018. SPIE

    2018

  9. [9]

    Exploring machine learning for semiconductor pro- cess optimization: A systematic review

    Chen, Ying-Lin, Sacchi, Sara, Dey, Bappaditya, Blanco, Victor, Halder, Sandip, Leray, Philippe and De Gendt, Ste- fan. “Exploring machine learning for semiconductor pro- cess optimization: A systematic review.”IEEE Transac- tions on Artificial IntelligenceVol. 5 No. 12 (2024): pp. 5969–5989

    2024

  10. [10]

    An approach for factory-wide control utilizing virtual metrol- ogy

    Khan,AftabA,Moyne,JamesRandTilbury,DawnM. “An approach for factory-wide control utilizing virtual metrol- ogy.”IEEE Transactions on semiconductor Manufacturing Vol. 20 No. 4 (2007): pp. 364–375

    2007

  11. [11]

    Developmentofthevirtualmetrologyforthenitridethick- ness in multi-layer plasma-enhanced chemical vapor depo- sition using plasma-information variables

    Roh, Hyun-Joon, Ryu, Sangwon, Jang, Yunchang, Kim, Nam-Kyun,Jin,Younggil,Park,SeolhyeandKim,Gon-Ho. “Developmentofthevirtualmetrologyforthenitridethick- ness in multi-layer plasma-enhanced chemical vapor depo- sition using plasma-information variables.”IEEE Trans- actions on Semiconductor ManufacturingVol. 31 No. 2 (2018): pp. 232–241

    2018

  12. [12]

    AI-Powered Next-Generation Technology for Semiconductor Optical Metrology: A Review

    Xu, Weiwang, Zhang, Houdao, Ji, Lingjing and Li, Zhongyu. “AI-Powered Next-Generation Technology for Semiconductor Optical Metrology: A Review.”Microma- chinesVol. 16 No. 8 (2025). DOI 10.3390/mi16080838

    work page doi:10.3390/mi16080838 2025

  13. [13]

    El-Kareh,BadihandHutter,LouN.Fundamentalsofsemi- conductorprocessingtechnology.SpringerScience&Busi- ness Media (2012)

    2012

  14. [14]

    Decision-based virtual metrol- ogyforadvancedprocesscontroltoempowersmartproduc- tion and an empirical study for semiconductor manufactur- ing

    Chien, Chen-Fu, Hung, Wei-Tse, Pan, Chin-Wei and Van Nguyen, Tran Hong. “Decision-based virtual metrol- ogyforadvancedprocesscontroltoempowersmartproduc- tion and an empirical study for semiconductor manufactur- ing.”Computers&IndustrialEngineeringVol.169(2022): p. 108245

    2022

  15. [15]

    Critical- dimension metrology and the scanning electron micro- scope

    Postek, Michael T and Vladár, András E. “Critical- dimension metrology and the scanning electron micro- scope.”Handbook of Silicon Semiconductor Metrology. CRC Press (2001): pp. 244–275

    2001

  16. [16]

    Spectro- scopicellipsometry: advancements,applicationsandfuture prospects in optical characterization

    Politano, Grazia Giuseppina and Versace, Carlo. “Spectro- scopicellipsometry: advancements,applicationsandfuture prospects in optical characterization.”Spectroscopy Jour- nalVol. 1 No. 3 (2023): pp. 163–181

    2023

  17. [17]

    Metrology

    Bunday, Benjamin and Orji, George. “Metrology.”2021 IEEEInternationalRoadmapforDevicesandSystemsOut- briefs: pp. 01–68. 2021. IEEE

    2021

  18. [18]

    Data-driven modeling in metrology–A short in- troduction, current developments and future perspectives

    Schneider, Linda-Sophie, Krauss, Patrick, Schiering, Na- dine, Syben, Christopher, Schielein, Richard and Maier, Andreas. “Data-driven modeling in metrology–A short in- troduction, current developments and future perspectives.” tm-TechnischesMessenVol.91No.9(2024): pp.480–503

    2024

  19. [19]

    Virtual metrology for semiconductor manufacturing applications

    Bertorelle, Nicola. “Virtual metrology for semiconductor manufacturing applications.”

  20. [20]

    Review of Ap- plications of Regression and Predictive Modeling in Wafer Manufacturing

    Chen, Hsuan-Yu and Chen, Chiachung. “Review of Ap- plications of Regression and Predictive Modeling in Wafer Manufacturing.”ElectronicsVol. 14 No. 20 (2025): p. 4083

    2025

  21. [21]

    CNN–BiLSTM–Attention-Based Hybrid- Driven Modeling for Diameter Prediction of Czochralski SiliconSingleCrystals

    Zhang, Pengju, Pan, Hao, Chen, Chen, Jing, Yiming and Liu, Ding. “CNN–BiLSTM–Attention-Based Hybrid- Driven Modeling for Diameter Prediction of Czochralski SiliconSingleCrystals.”Crystals(2073-4352)Vol.16No.1 (2026)

    2073

  22. [22]

    AMTransformer: A Koopman theory-based transformer for learning additive manufacturing dynamics in laser processes

    Lee, Suk Ki and Ko, Hyunwoong. “AMTransformer: A Koopman theory-based transformer for learning additive manufacturing dynamics in laser processes.”International JournalofAIforMaterialsandDesignVol.1No.2(2024): pp. 76–91

    2024

  23. [23]

    Virtual Metrology Based on Graph Convolu- tional Neural Network for Semiconductor PVD Process

    Zhou,Longfei,Jin,Dong,Chen,Shuangwu,Yang,Jianand Xie, Jian. “Virtual Metrology Based on Graph Convolu- tional Neural Network for Semiconductor PVD Process.” 20245thInternationalConferenceonArtificialIntelligence 9 andElectromechanicalAutomation(AIEA):pp.1053–1058

  24. [24]

    Graphattention-baseddynamicalandcausalspatiotempo- ral learning for anomaly detection in additive manufactur- ing

    Lee, Suk Ki, Kim, Wonah, Lee, Sungbeom, Park, Jeonghyeon, Chun, Sejin, Yeung, Ho and Ko, Hyunwoong. “Graphattention-baseddynamicalandcausalspatiotempo- ral learning for anomaly detection in additive manufactur- ing.”VirtualandPhysicalPrototypingVol.21No.1(2026): p. e2611194. DOI 10.1080/17452759.2025.2611194

    work page doi:10.1080/17452759.2025.2611194 2026

  25. [25]

    Multi-stage pro- cess diagnosis networks in semiconductor manufacturing

    Choi, Jongwon and Kim, Seoung Bum. “Multi-stage pro- cess diagnosis networks in semiconductor manufacturing.” IEEE AccessVol. 12 (2024): pp. 39495–39504

    2024

  26. [26]

    XplAInable: Explainable AI Smoke Detection at the Edge

    Lehnert, Alexander, Gawantka, Falko, During, Jonas, Just, Franz and Reichenbach, Marc. “XplAInable: Explainable AI Smoke Detection at the Edge.”Big Data and Cognitive ComputingVol. 8 No. 5 (2024): p. 50

    2024

  27. [27]

    Seshan,Krishna.Handbookofthinfilmdeposition.William Andrew (2012). 10

    2012