pith. sign in

arxiv: 2504.03990 · v2 · pith:FV5VZH77new · submitted 2025-04-04 · 🧮 math.NA · cs.NA· physics.comp-ph

Parametric Operator Inference to Simulate the Purging Process in Semiconductor Manufacturing

Seunghyon Kang , Hyeonghun Kim , Boris Kramer This is my paper

Pith reviewed 2026-05-22 20:45 UTC · model grok-4.3

classification 🧮 math.NA cs.NAphysics.comp-ph
keywords reduced-order modelingoperator inferenceparametric modelingsemiconductor manufacturingpurging processcomputational fluid dynamicsPECVD
0
0 comments X

The pith

Parametric operator inference creates reduced-order models that predict purging flow fields in semiconductor chambers across 25 parameter combinations with a maximum error of 9.32%.

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

The paper applies parametric Operator Inference to learn reduced-order models from CFD data for the argon purging process inside a PECVD chamber. It trains nine separate ROMs on varying inlet mass flow rates and outlet pressures, then uses linear interpolation to obtain predictions at 25 total parameter points including 16 combinations absent from the training set. The resulting models are evaluated on 64% of the data and shown to reproduce the key flow features while delivering a 142-fold reduction in online compute time relative to the full CFD solver. A reader would care because such models could support rapid evaluation of contamination-control strategies during semiconductor manufacturing without repeated expensive simulations.

Core claim

The parametric OpInf framework learns nine ROMs from CFD snapshot data for different argon mass flow rates at the inlet and different outlet pressures. It then interpolates these ROMs to forecast the flow field for 25 parameter combinations, 16 of which are unseen during training. Trained on only 36% of the available data, the interpolated models reproduce the purging behavior across the full parameter domain with a maximum error of 9.32% and run approximately 142 times faster than the original CFD simulation.

What carries the argument

Parametric Operator Inference (OpInf), a non-intrusive data-driven method that learns low-dimensional dynamical models from simulation snapshots and interpolates the learned operators across parameter values.

Load-bearing premise

The simplified CFD model that omits plasma dynamics and chemical reactions still captures the essential purging flow features, and linear interpolation among the nine learned ROMs accurately represents the flow at the 16 unseen parameter points.

What would settle it

Compute a full-order CFD solution at one of the 16 unseen parameter combinations and measure whether the L2 or pointwise error between that solution and the interpolated OpInf ROM exceeds 9.32%.

Figures

Figures reproduced from arXiv: 2504.03990 by Boris Kramer, Hyeonghun Kim, Seunghyon Kang.

Figure 1
Figure 1. Figure 1: Illustration of the impact of particle contamination on subsequent wafer manufacturing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Computational domain of a simplified PECVD chamber. B.C.1: mass flow inlet boundary [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Left: velocity field, Right: streamlines at the onset of the purging process, initiated upon [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Normalized singular value decay. (b) the cumulative energy [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a) Datasets used for training and testing. Of the 25 datasets, nine (36%) are used for [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Contour plots of the temperature in the vertical cross-section for the case with the largest [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Contour plots of the y-velocity vy in the vertical cross-section for the case with the largest error, (µp, µq) = (0.56, 0.95). Each column represents t = 0s, 0.5s, and 1s. (a) FOM, (b) ROM predictions, (c) Pointwise normalized absolute error. We compare the y-velocity and radial velocity flow fields at the monitor location for the case with the largest error, (µp, µq) = (0.56, 0.95), in Figures 10 and 11. … view at source ↗
Figure 8
Figure 8. Figure 8: Contour plots of the radial velocity vr = p v 2 x + v 2 z in the vertical cross-section for the case with the largest error, (µp, µq) = (0.56, 0.95). Each column represents t = 0s, 0.5s, and 1s. (a) FOM, (b) ROM predictions, (c) Pointwise normalized absolute error [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Monitor location: a quarter circular xz-plane with a 150 mm radius, positioned 1 mm above the wafer surface. (a) (b) (c) [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Contour plots of the y-velocity at the monitor location for the case with the largest error, [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Contour plots of the radial velocity at the monitor location for the case with the largest [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
read the original abstract

This work presents the application of parametric Operator Inference (OpInf) -- a nonintrusive reduced-order modeling (ROM) technique that learns a low-dimensional representation of a high-fidelity model -- to the numerical model of the purging process in semiconductor manufacturing. Leveraging the data-driven nature of the OpInf framework, we aim to forecast the flow field within a plasma-enhanced chemical vapor deposition (PECVD) chamber using computational fluid dynamics (CFD) simulation data. Our model simplifies the system by excluding plasma dynamics and chemical reactions, while still capturing the key features of the purging flow behavior. The parametric OpInf framework learns nine ROMs based on varying argon mass flow rates at the inlet and different outlet pressures. It then interpolates these ROMs to predict the system's behavior for 25 parameter combinations, including 16 scenarios that are not seen in training. The parametric OpInf ROMs, trained on 36\% of the data and tested on 64\%, demonstrate accuracy across the entire parameter domain, with a maximum error of 9.32\%. Furthermore, the ROM achieves an approximate 142-fold speedup in online computations compared to the full-order model CFD simulation. These OpInf ROMs may be used for fast and accurate predictions of the purging flow in the PECVD chamber, which could facilitate effective particle contamination control 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 / 2 minor

Summary. The manuscript applies parametric Operator Inference (OpInf) to create reduced-order models of the purging flow in a PECVD chamber for semiconductor manufacturing. Nine ROMs are learned from CFD data at selected combinations of argon inlet mass flow rate and outlet pressure; these ROMs are then interpolated to predict the flow field at 25 parameter points (16 of which are unseen in training). The underlying CFD model omits plasma dynamics and chemical reactions. Reported results include a maximum error of 9.32% on the 64% test portion of the data and an online speedup of approximately 142 times relative to the full-order CFD simulation.

Significance. If the accuracy claims hold under the stated modeling assumptions, the work supplies a practical, nonintrusive surrogate that could support rapid evaluation of purging scenarios for contamination control. The parametric OpInf construction and the reported computational speedup constitute a concrete engineering application of data-driven ROM techniques.

major comments (2)
  1. [Abstract] Abstract: the central claim that linear interpolation of the nine learned reduced operators produces accurate predictions at the 16 unseen parameter combinations (with max error 9.32%) is load-bearing. No evidence is supplied that the map from (argon flow rate, outlet pressure) to the reduced operators is sufficiently smooth for linear interpolation, nor is cross-validation on a denser parameter grid or an alternative interpolation scheme reported.
  2. [Abstract] Abstract: the modeling assumption that excluding plasma dynamics and chemical reactions leaves the dominant purging flow features intact for contamination-control purposes is not accompanied by a quantitative sensitivity study or comparison against a model that retains those effects.
minor comments (2)
  1. The abstract states a 36%/64% training/testing split but does not specify how the split was performed across the 25-point parameter grid or which norm is used to compute the 9.32% maximum error.
  2. Clarify whether the interpolation is performed on the reduced operators themselves or on their coefficients, and state the precise interpolation procedure (e.g., linear in parameter space).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below with the strongest honest defense of the manuscript while acknowledging where revisions can strengthen the presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that linear interpolation of the nine learned reduced operators produces accurate predictions at the 16 unseen parameter combinations (with max error 9.32%) is load-bearing. No evidence is supplied that the map from (argon flow rate, outlet pressure) to the reduced operators is sufficiently smooth for linear interpolation, nor is cross-validation on a denser parameter grid or an alternative interpolation scheme reported.

    Authors: The reported maximum error of 9.32% on the 16 unseen test points constitutes empirical evidence that linear interpolation is effective for this parameter regime and sampling density. The manuscript does not supply a theoretical smoothness analysis or results from a denser grid or alternative schemes, as generating additional high-fidelity CFD data is computationally expensive. We will revise the methods and results sections to explicitly discuss the parameter sampling strategy and to present the observed interpolation accuracy as practical validation of the approach. revision: partial

  2. Referee: [Abstract] Abstract: the modeling assumption that excluding plasma dynamics and chemical reactions leaves the dominant purging flow features intact for contamination-control purposes is not accompanied by a quantitative sensitivity study or comparison against a model that retains those effects.

    Authors: The manuscript states the modeling simplification explicitly and justifies it on the basis that the purging flow for particle contamination control is dominated by the inert-gas transport captured in the CFD model. No quantitative sensitivity study against a plasma-inclusive multiphysics model is provided, as constructing and validating such a model lies outside the scope of the present work on parametric OpInf for flow data. We will add a concise limitations paragraph in the introduction or conclusions to clarify this assumption and its implications. revision: yes

Circularity Check

0 steps flagged

No significant circularity: data-driven parametric OpInf from external CFD simulations

full rationale

The paper trains parametric OpInf reduced-order models on CFD data at 9 parameter combinations (36% of data) and interpolates the learned operators to predict at 25 total points including 16 unseen ones (64% held-out test data). This is a standard non-intrusive data-driven workflow with explicit validation against full-order simulations; no derivation step reduces by construction to its own fitted inputs, no load-bearing self-citation chains, and no ansatz or uniqueness claim that loops back to the target result. The reported 9.32% max error and 142x speedup are empirical outcomes on external data rather than tautological predictions.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

Review performed on abstract only; ledger entries are inferred from the high-level description of the OpInf workflow and the stated modeling simplifications.

free parameters (2)
  • ROM dimension / rank
    OpInf requires choosing the reduced dimension; abstract does not specify how it was selected or whether it was fitted.
  • Interpolation weights or kernel parameters
    Parametric interpolation among the nine base ROMs requires additional choices whose values are not reported.
axioms (2)
  • domain assumption The flow dynamics admit a low-dimensional linear operator representation learnable from snapshot data
    Core premise of the Operator Inference framework invoked when the abstract states that nine ROMs are learned from CFD data.
  • ad hoc to paper Excluding plasma and chemical reactions does not alter the dominant flow features relevant to purging
    Explicit modeling choice stated in the abstract that underpins the entire reduced model.

pith-pipeline@v0.9.0 · 5775 in / 1495 out tokens · 73505 ms · 2026-05-22T20:45:08.794337+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

  1. [1]

    Behavior of particle reflected by turbo molecular pump in plasma etching apparatus,

    H. Kobayashi, K. Maeda, and M. Izawa, “Behavior of particle reflected by turbo molecular pump in plasma etching apparatus,”IEEE Transactions on Semiconductor Manufacturing, vol. 22, pp. 462–467, 2009

    work page 2009

  2. [2]

    Parti- cle reduction and control in plasma etching equipment,

    T. Moriya, H. Nakayama, H. Nagaike, Y. Kobayashi, M. Shimada, and K. Okuyama, “Parti- cle reduction and control in plasma etching equipment,”IEEE Transactions on Semiconductor Manufacturing, vol. 18, pp. 477–486, 2005

    work page 2005

  3. [3]

    A particle reduction strategy for plasma etching process,

    J. Jeong, Y. Kim, J. Lee, and Y. Kim, “A particle reduction strategy for plasma etching process,” 2024 35th Annual SEMI Advanced Semiconductor Manufacturing Conference (ASMC), pp. 1–4, 2024

    work page 2024

  4. [4]

    Particle formation and its control in dual frequency plasma etching reactors,

    M. Kim, H.-W. Cheong, and K.-W. Whang, “Particle formation and its control in dual frequency plasma etching reactors,”Journal of Vacuum Science and Technology A, vol. 33, p. 041303, 2015

    work page 2015

  5. [5]

    Kernel-density-based particle defect management for semi- conductor manufacturing facilities,

    S. H. Park, S. Kim, and J.-G. Baek, “Kernel-density-based particle defect management for semi- conductor manufacturing facilities,”Applied Sciences, vol. 8, no. 2, p. 224, 2018

    work page 2018

  6. [6]

    Backside defect monitoring strategy and improvement in the advanced semiconductor manufacturing,

    J. G. Zhou, H. Chen, Y. Long, K. Wang, H. Gua, and F. Liu, “Backside defect monitoring strategy and improvement in the advanced semiconductor manufacturing,”2021 China Semiconductor Technology International Conference, vol. 29, pp. 1–5, 2021

    work page 2021

  7. [7]

    Experimental and analytical study of submicrometer particle removal from deep trenches,

    K. Bakhtari, R. O. Guldiken, A. A. Busnaina, and J.-G. Park, “Experimental and analytical study of submicrometer particle removal from deep trenches,”Journal of The Electrochemical Society, vol. 153, no. 9, pp. C603–C607, 2006

    work page 2006

  8. [8]

    Particle cleaning technologies to meet advanced semiconductor device process requirements,

    H. F. Okorn-Schmidt, F. Holsteyns, A. Lippert, D. Mui, M. Kawaguchi, C. Lechner, P. E. Frommhold, T. Nowak, F. Reuter, and M. B. Pique, “Particle cleaning technologies to meet advanced semiconductor device process requirements,”ESC Journal of Solid State Science and Technology, vol. 3, no. 1, pp. N3069–N3080, 2013

    work page 2013

  9. [9]

    Computational fluid dynamics analysis of particle deposition induced by a showerhead electrode in a capacitively coupled plasma reactor,

    H. J. Kim and J. H. Yoon, “Computational fluid dynamics analysis of particle deposition induced by a showerhead electrode in a capacitively coupled plasma reactor,”Coatings, vol. 11, p. 1004, 2021

    work page 2021

  10. [10]

    Particle deposition velocity onto a wafer or a photomask in a laminar parallel flow,

    S.-J. Yook, H.-J. Hwang, K.-S. Lee, and K. ho Ahn, “Particle deposition velocity onto a wafer or a photomask in a laminar parallel flow,”Journal of The Electrochemical Society, vol. 157, pp. H692–H698, 2010

    work page 2010

  11. [11]

    A survey on machine and deep learning in semicon- ductor industry: methods, opportunities, and challenges,

    A. C. Huang, S. H. Meng, and T. J. Huang, “A survey on machine and deep learning in semicon- ductor industry: methods, opportunities, and challenges,”Cluster Computing, vol. 26, no. 6, pp. 3437–3472, 2023

    work page 2023

  12. [12]

    Exploring ma- chine learning for semiconductor process optimization: A systematic review,

    Y.-L. Chen, S. Sacchi, B. Dey, V. Blanco, S. Halder, P. Leray, and S. D. Gendt, “Exploring ma- chine learning for semiconductor process optimization: A systematic review,”IEEE Transactions on Artificial Intelligence, vol. 5, no. 12, pp. 5969–5989, 2024

    work page 2024

  13. [13]

    Machine learning-based modeling and operation for ALD of SiO2 thin-films using data from a multiscale CFD simulation,

    Y. Ding, Y. Zhang, Y. M. Ren, G. Orkoulas, and P. D. Christofides, “Machine learning-based modeling and operation for ALD of SiO2 thin-films using data from a multiscale CFD simulation,” Chemical Engineering Research and Design, vol. 151, pp. 131–145, 2019

    work page 2019

  14. [14]

    Machine learning-based modeling and operation of plasma-enhanced atomic layer deposition of hafnium oxide thin films,

    Y. Ding, Y. Zhang, H. Y. Chung, and P. D. Christofides, “Machine learning-based modeling and operation of plasma-enhanced atomic layer deposition of hafnium oxide thin films,”Computers & Chemical Engineering, vol. 144, p. 107148, 2021. 15

    work page 2021

  15. [15]

    Machine learning enabled optimization of showerhead design for semiconductor deposition process,

    Z. Jin, D. D. Lim, X. Zhao, M. Mamunuru, S. Roham, and G. X. Gu, “Machine learning enabled optimization of showerhead design for semiconductor deposition process,”Journal of Intelligent Manufacturing, vol. 35, no. 2, pp. 925–935, 2024

    work page 2024

  16. [16]

    High-speed pre- diction of computational fluid dynamics simulation in crystal growth,

    Y. Tsunooka, N. Kokubo, G. Hatasa, S. Harada, M. Tagawa, and T. Ujihara, “High-speed pre- diction of computational fluid dynamics simulation in crystal growth,”CrystEngComm, vol. 20, no. 41, pp. 6546–6550, 2018

    work page 2018

  17. [17]

    Detection and identification of particles on silicon wafers based on light scattering and absorption spectroscopy and machine learning,

    F. Zhoua, X. Fua, S. Chena, and M. B. G. Jun, “Detection and identification of particles on silicon wafers based on light scattering and absorption spectroscopy and machine learning,”Manufac- turing Letters, vol. 35, pp. 991–998, 2023

    work page 2023

  18. [18]

    Data-driven operator inference for nonintrusive projection-based model reduction,

    B. Peherstorfer and K. Willcox, “Data-driven operator inference for nonintrusive projection-based model reduction,”Computer Methods in Applied Mechanics and Engineering, vol. 306, pp. 196– 215, 2016

    work page 2016

  19. [19]

    Learning nonlinear reduced models from data with operator inference,

    B. Kramer, B. Peherstorfer, and K. E. Willcox, “Learning nonlinear reduced models from data with operator inference,”Annual Review of Fluid Mechanics, vol. 56, no. 1, pp. 521–548, 2024

    work page 2024

  20. [20]

    Nonintrusive reduced-order models for parametric partial differential equations via data-driven operator inference,

    S. A. McQuarrie, P. Khodabakhshi, and K. E. Willcox, “Nonintrusive reduced-order models for parametric partial differential equations via data-driven operator inference,”SIAM Journal on Scientific Computing, vol. 45, no. 4, pp. A1917–A1946, 2023

    work page 2023

  21. [21]

    Learning reduced-order dynamics for parametrized shallow water equations from data,

    S. Yıldız, P. Goyal, P. Benner, and B. Karasözen, “Learning reduced-order dynamics for parametrized shallow water equations from data,”International Journal for Numerical Meth- ods in Fluids, vol. 93, no. 8, pp. 2803–2821, 2021

    work page 2021

  22. [22]

    Parametric non-intrusive reduced-order models via operator inference for large-scale rotating detonation engine simulations,

    I. Farcas, R. Gundevia, R. Munipalli, and K. E. Willcox, “Parametric non-intrusive reduced-order models via operator inference for large-scale rotating detonation engine simulations,” inAIAA SCITECH 2023 Forum, 2023, p. 0172

    work page 2023

  23. [23]

    Data-driven reduced-order models via regularised operator inference for a single-injector combustion process,

    S. A. McQuarrie, C. Huang, and K. E. Willcox, “Data-driven reduced-order models via regularised operator inference for a single-injector combustion process,”Journal of the Royal Society of New Zealand, vol. 51, no. 2, p. 194–211, 2021

    work page 2021

  24. [24]

    Learning physics-based reduced-order mod- els for a single-injector combustion process,

    R. Swischuk, B. Kramer, C. Huang, and K. Willcox, “Learning physics-based reduced-order mod- els for a single-injector combustion process,”AIAA Journal, vol. 58, no. 6, p. 2658–2672, Jun. 2020

    work page 2020

  25. [25]

    Reduced operator inference for nonlinear partial dif- ferential equations,

    E. Qian, I.-G. Farcas, and K. Willcox, “Reduced operator inference for nonlinear partial dif- ferential equations,”SIAM Journal on Scientific Computing, vol. 44, no. 4, pp. A1934–A1959, 2022

    work page 2022

  26. [26]

    Data-driven model reduction via operator inference for coupled aeroelastic flutter,

    B. G. Zastrow, A. Chaudhuri, K. Willcox, A. S. Ashley, and M. C. Henson, “Data-driven model reduction via operator inference for coupled aeroelastic flutter,” inAIAA Scitech 2023 Forum, 2023, p. 0330

    work page 2023

  27. [27]

    Operator inference and physics-informed learn- ing of low-dimensional models for incompressible flows,

    P. Benner, P. Goyal, J. Heiland, and I. P. Duff, “Operator inference and physics-informed learn- ing of low-dimensional models for incompressible flows,”Electronic Transactions on Numerical Analysis, vol. 56, pp. 28–51, 2022

    work page 2022

  28. [28]

    Reduced- order modeling of the two-dimensional Rayleigh–Bénard convection flow through a non-intrusive operator inference,

    P. R. B. Rocha, J. L. de Sousa Almeida, M. S. de Paula Gomes, and A. C. N. Junior, “Reduced- order modeling of the two-dimensional Rayleigh–Bénard convection flow through a non-intrusive operator inference,”Engineering Applications of Artificial Intelligence, vol. 126, p. 106923, 2023. 16

    work page 2023

  29. [29]

    Sampling low-dimensional markovian dynamics for preasymptotically recovering reduced models from data with operator inference,

    B. Peherstorfer, “Sampling low-dimensional markovian dynamics for preasymptotically recovering reduced models from data with operator inference,”SIAM Journal on Scientific Computing, vol. 42, no. 5, pp. A3489–A3515, 2020

    work page 2020

  30. [30]

    The quickhull algorithm for convex hulls,

    C. B. Barber, D. P. Dobkin, and H. Huhdanpaa, “The quickhull algorithm for convex hulls,”ACM Transactions on Mathematical Software, vol. 22, no. 4, p. 469–483, 1996

    work page 1996

  31. [31]

    SciPy 1.0: fundamental algorithms for scientific computing in Python,

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Brightet al., “SciPy 1.0: fundamental algorithms for scientific computing in Python,”Nature methods, vol. 17, no. 3, pp. 261–272, 2020

    work page 2020

  32. [32]

    Operator Inference package,

    “Operator Inference package,” https://github.com/Willcox-Research-Group/ rom-operator-inference-Python3, accessed: 2025-03-04

    work page 2025

  33. [33]

    Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions,

    N. Halko, P.-G. Martinsson, and J. A. Tropp, “Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions,”SIAM Review, vol. 53, no. 2, pp. 217–288, 2011

    work page 2011

  34. [34]

    Scikit-learn: Ma- chine learning in python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, A. Müller, J. Nothman, G. Louppe, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and Édouard Duchesnay, “Scikit-learn: Ma- chine learning in python,” 2018. 17

    work page 2018