Physics-Informed Neural Operator for Warm-Starting Background-Decomposed and Preconditioned PSFD: Enabling Scalable 3-D EUV Mask Simulation

Doyun Kim; Werner Gillijns

REVIEW 4 major objections 6 minor 35 references

This paper shows that a Fourier neural operator trained only on the scattering equations, with no precomputed electromagnetic solutions, reaches about 7e-3 accuracy on held-out EUV masks and cuts a rigorous solver's end-to-end time from abo

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 02:44 UTC pith:Y5P2ZO77

load-bearing objection A credible warm-start PINO for EUV mask simulation; the missing smoothing width and other reporting gaps leave the physical claim under-supported. the 4 major comments →

arxiv 2607.25330 v1 pith:Y5P2ZO77 submitted 2026-07-28 physics.optics cs.AIcs.LGphysics.app-ph

Physics-Informed Neural Operator for Warm-Starting Background-Decomposed and Preconditioned PSFD: Enabling Scalable 3-D EUV Mask Simulation

Doyun Kim , Werner Gillijns This is my paper

classification physics.optics cs.AIcs.LGphysics.app-ph

keywords EUV lithographyneural operatorphysics-informed trainingelectromagnetic scatteringpseudo-spectral frequency-domainwarm-startmask 3D effectsFourier neural operator

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's central claim is that a physics-informed neural operator can replace the usual training-data bottleneck for EUV mask scattering: instead of learning from precomputed fields, it minimizes the residual of the pseudo-spectral frequency-domain (PSFD) Maxwell equations, with the thick multilayer substrate folded into an analytic reflection operator. If this works, mask near-field prediction no longer needs a labeled dataset of rigorous solutions, and the same surrogate can warm-start the rigorous solver. The paper reports a mean absolute error of about 7e-3 in scattered intensity on held-out patterns at native resolution, and about 3e-2 when transferring to a four-times-finer lateral grid without retraining. Combined with spectral damping, the warm start reaches a practically useful residual in about 75 iterations (about 2 minutes) instead of about 1,000 iterations (about 30 minutes). A sympathetic reader would care because mask 3D effects are a main accuracy bottleneck in computational lithography, and rigorous solvers are too slow for large-scale use.

Core claim

The central discovery, as the authors state it, is that a Fourier neural operator can be trained self-consistently with the background-decomposed PSFD formulation, retaining full-vector coupling between the mask and the multilayer response without invoking a finite-order Born approximation and without any precomputed EM field solutions. The loss is the PSFD residual plus a frequency-weighted penalty, evaluated on randomly sampled mask designs each training iteration. The trained surrogate predicts the three-dimensional vectorial scattered field; the reported accuracy is a mean absolute error of about 7e-3 in scattered intensity on 100 held-out masks, rising to about 3e-2 on a four-times-fine

What carries the argument

The load-bearing machinery is an anisotropically factorized Fourier neural operator, in which each layer applies 2-D Fourier filters on lateral slices and a 1-D Fourier filter along the axial direction, followed by pointwise channel mixing. It is trained to minimize a physics-informed loss built from the background-decomposed PSFD equation. Background decomposition shrinks the computational domain to the absorber region; the thick multilayer's response is applied analytically through a transfer-matrix-based reflection operator at every iteration, avoiding the grid-dependent phase error and memory cost of explicitly meshing the roughly 280-nm stack. A scalar spectral-damping right preconditio

Load-bearing premise

The load-bearing premise is that the Gaussian-smoothed binary mask geometry, used for both the PINO input and the PSFD reference, faithfully represents the physical EUV mask; the paper does not report the smoothing width or quantify how it changes the scattered field.

What would settle it

Compare the reported mean absolute error and warm-start iteration counts against a reference with much smaller smoothing, or ideally against the original sharp mask edges; if the error rises well above 7e-3 or the 75-iteration figure degrades, the smoothing is doing the accuracy work and the practical claim does not transfer.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Since training uses the governing equations rather than precomputed solutions, the approach scales to new mask design spaces without paying for a large rigorous-simulation dataset.
The warm-start plus damping configuration reaches the practical 5e-2 residual level in about 75 iterations, a roughly 13x reduction over the cold unpreconditioned solve and about 2x over damping alone, with end-to-end time dropping from about 30 minutes to about 2 minutes.
The trained surrogate transfers to a four-times-finer lateral grid without retraining, at a mean absolute error of about 3e-2, with errors concentrated at material interfaces.
The background decomposition isolates the multilayer response analytically, so surrogate errors are not entangled with numerical phase errors from discretizing the stack.
At tight tolerances (relative residual 1e-7), the warm-start offers little asymptotic advantage: both preconditioned configurations require roughly 2,000 iterations, so the benefit is confined to the early practical-convergence regime.

Where Pith is reading between the lines

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

The paper does not report the Gaussian smoothing width or its lithographic effect; an important extension is to test the same pipeline on unsmoothed or sharp-edge masks against a rigorously converged reference, since the reported accuracy is measured against a smoothed model.
The threshold-contour metric divided by four is a near-field proxy; end-to-end validation through aerial-image or resist models would be needed to know whether the 75-iteration operating point actually holds for printed CD control.
The factorized lateral-axial architecture and analytic reflection operator are likely compatible with oblique illumination and partially coherent imaging, which could let the warm-start framework extend into the full optical proximity correction flow.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A credible warm-start PINO for EUV mask simulation; the missing smoothing width and other reporting gaps leave the physical claim under-supported.

read the letter

This is a credible, practical paper from imec: they train a factorized Fourier neural operator on the PSFD residual (no precomputed EM fields) and use it to warm-start their preconditioned EUV mask solver. The new combination works: on their test set, warm-starting cuts the iteration count to reach their practical threshold from ~1000 to ~75, and the wall-clock drop (30 min to 2 min) is reported including overhead. The training procedure is self-consistent and the convergence curves are the most convincing part.

What the paper does well: the background-decomposed PSFD operator with the TMM reflection operator is a sensible physics-informed loss; the high-frequency weighting is reasonable; and they are honest that the warm-start advantage mostly disappears when you demand a tight residual (1e-7). That kind of reporting earns trust.

The soft spots are real but not disqualifying. The main worry—shared by your stress-test note—is the Gaussian smoothing. They smooth the binary mask geometry, and they apply the same smoothing to the PSFD reference, so the reported MAE is against a smoothed model. The smoothing width is never reported. If the sigma is large enough to smear absorber edges, the PINO may be learning a surrogate for a model that does not reproduce the true EUV near-field. They cite standard subpixel-smoothing practice, but they don't test the sensitivity to sigma or compare to an unsmoothed rigorous solver. That's a missing experiment, and it is load-bearing for the physical claim.

Second, the MAE is an absolute number with no typical intensity scale, so '7e-3' is hard to interpret. Third, there is a clear domain-size contradiction: the training domain is described as 600x600x200 nm^3 at 4-nm spacing, then later as 1200x1200x200 for the same spacing. It's an obvious typo, but it signals sloppy proofreading. Fourth, most architecture and training hyperparameters—L, d_v, d_h, k_max, loss weights, Adam settings, and the smoothing width—are missing. For a paper whose contribution is partly 'we engineered this to work,' that's too much omitted.

None of this invalidates the central result. The residual-based training is genuinely data-free, and the warm-start acceleration is demonstrated by direct comparison. This deserves a serious referee. I would ask the authors to report the missing numbers, fix the typo, and add at least one check of the smoothing sensitivity, ideally against an independent solver on the true binary geometry. If they supply that, the paper becomes a solid contribution for computational lithography.

I would send it to peer review rather than desk reject.

Referee Report

4 major / 6 minor

Summary. This paper proposes a physics-informed neural operator (PINO) for 3-D EUV mask scattering. The network is a Fourier neural operator factorized into a lateral 2-D branch and an axial 1-D branch, and it is trained by minimizing the residual of a background-decomposed pseudo-spectral frequency-domain (PSFD) equation; no precomputed EM field solutions are used as training labels. The mask geometry is Gaussian-smoothed and supplied as the network input. The paper reports an MAE of about 7e-3 for scattered intensity on a representative held-out mask at training resolution, an MAE of about 3e-2 when transferring to a four-times-finer lateral grid, and a warm-starting scheme that, combined with a spectral-damping preconditioner, reaches a practical convergence criterion in about 75 iterations versus about 1,000 for the unpreconditioned solver, corresponding to roughly 2 minutes versus 30 minutes wall-clock.

Significance. If the reported accuracy and transferability hold, the paper makes a significant contribution: a residual-based, label-free neural operator for full-vector EUV mask scattering. The factorized architecture addresses a real 3-D scalability bottleneck, and the background decomposition avoids both the Born approximation and multilayer discretization errors. The controlled comparison among null, preconditioned, and warm-started solvers, together with a threshold-width metric averaged over 100 test patterns, is a strength. The main uncertainties are the smoothed-geometry assumption and the lack of aggregate error statistics, which currently prevent a firm assessment of the central accuracy claims.

major comments (4)

[Section 3, PINO surrogate paragraph; Eq. (5)] The central accuracy claim is defined with respect to a PSFD reference that uses the same Gaussian smoothing as the PINO input. The smoothing width is not reported, and no comparison is made to an unsmoothed geometry or an independent rigorous solver (e.g., RCWA/FDTD). Since EUV mask scattering is strongly edge- and phase-sensitive, the reported 7e-3 MAE and the warm-start benefit may describe only a regularized surrogate problem, not the physical binary-mask problem. Please report the smoothing width, quantify its effect on the threshold-width/CD metric, and provide at least one comparison against an unsmoothed reference or explicitly state that the contribution is to the smoothed model.
[Section 3, PINO surrogate and Application to finer inference grids; Figs. 2 and 4] The MAE values 7e-3 and 3e-2 are unnormalized absolute errors, and they appear to be reported for single representative masks (Fig. 2 and Fig. 3-4) despite the stated evaluation on 100 held-out masks. Without per-mask statistics and a normalization (e.g., relative to the mean |E_y|^2 or the reference intensity range), the magnitude of 7e-3 cannot be interpreted, and the finer-grid claim of 3e-2 is not statistically supported. Please report the mean, standard deviation, and quantiles over all held-out masks at both resolutions.
[Section 2 and Section 3] For reproducibility and independent verification, the manuscript must report the numerical settings that define the method: Gaussian smoothing width; lambda_hf, w0, eta_hf in Eqs. (8)-(9); L, d_v, d_h, k_max in Eqs. (2)-(4) and the total parameter count; learning rate, batch size, and number of training iterations; m_min, k_ref, gamma, epsilon in Eq. (13); and I_th, w in Eq. (17). Also, the physical domain size is stated inconsistently: Section 3 gives 600x600x200 nm^3 for the training samples, while the finer-grid section says the same 300x300x100 grid at 4 nm spacing corresponds to 1200x1200x200 nm^3. This must be corrected.
[Section 3, Application to finer inference grids; Figs. 3-4] The transfer-to-finer-grid demonstration is performed on a single mask, yet the conclusion generalizes this result. Since this transferability is one of the paper's central claims, please provide an aggregate evaluation over the held-out test set or explicitly restrict the claim. At minimum, report error statistics over a small set of masks and state whether the 3e-2 value is a mean or a single-case result.

minor comments (6)

[Figure 4 caption and text] The text states that Fig. 4(a) is the PINO inference and Fig. 4(b) is the PSFD reference, while the caption assigns the opposite roles. Please correct this inconsistency.
[Eqs. (9) and (17)] The symbol w is used both for the frequency-dependent weight in Eq. (9) and for the sigmoid transition width in Eq. (17). Please rename one of them to avoid confusion.
[Section 3, paragraph after Fig. 4] The phrase 'roughly half an order of magnitude larger' is inaccurate: 3e-2/7e-3 is about a factor of 4.3, which is closer to 0.6 of a decade. Suggest 'a factor of about 4'.
[Figure 3 caption] Typo: 'gemetry' should be 'geometry'.
[Section 3, PINO warm-start paragraph] Please state explicitly whether the convergence curves in Fig. 5 and the wall-clock timings are computed on the coarse training grid or the finer 1 nm grid, and whether the PINO warm start at the finer grid uses interpolation or continuous-frequency kernel evaluation. This is needed to interpret the timing comparison.
[Section 3, PINO surrogate paragraph] Please report the training wall-clock time and the number of network parameters, so the 2-minute inference/warm-start time can be contextualized.

Circularity Check

0 steps flagged

No construction-level circularity: the PINO is trained on the PSFD residual and evaluated against an independently computed PSFD reference, so the accuracy and warm-start claims do not reduce to fitted inputs.

full rationale

The derivation chain is not circular. The PINO loss (Eq. 8) is the residual of the vectorial PSFD/Helmholtz operator, and the paper explicitly states that no precomputed EM field solutions are used for training. Accuracy is assessed by comparing the trained network to a separately converged background-decomposed PSFD solve on held-out mask patterns (Figs. 2 and 4). The warm-start claim is measured by the iteration count of the same Krylov solver initialized with the PINO field (Eqs. 15-16, Fig. 5); this is an empirical property of the combined system, not a quantity fitted into the loss. The shared Gaussian smoothing of the PINO input and the PSFD reference is a modeling assumption that defines the problem being solved, not a circular step; the paper consistently labels the target as the reference PSFD solution. The main dependency on the authors' prior work (Ref. [24]) supplies the discretized operator and preconditioner, but this is normal methodological self-citation rather than a circular reduction: the operator is Maxwell's equations discretized pseudospectrally, and the paper includes independent checks against analytic TMM and against a standard full-domain PSFD solver for the multilayer benchmark (Fig. 1). Any concern about the absence of an independent rigorous solver for full 3D masks is a correctness/validation gap, not a circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The free parameters are mostly unreported hyperparameters and manually set thresholds that the central accuracy/convergence numbers depend on; their omission is the main reproducibility gap. The axioms are standard physics/ML assumptions, with the strongest unverified one being the physical fidelity of the Gaussian-smoothed mask model.

free parameters (5)

Gaussian smoothing width sigma = not reported
Applied to the binary mask indicator before network input and PSFD material distribution; width affects spectral content and Gibbs ringing, but is not given.
High-frequency loss weights lambda_hf, w_0, eta_hf = not reported
Weights in Eqs. (8)-(9); they tune the residual and spectral penalty but are not reported.
Architecture hyperparameters (L, d_v, d_h, k_max, LR, batch size) = not reported
Number of layers, channels, Fourier mode truncations, optimizer settings are needed to reproduce the method but are absent.
Preconditioner parameters (m_min, k_ref, gamma, epsilon) = not reported
Defined in Eq. (13); the values used in the Figure 5 experiments are not stated.
Threshold metric parameters I_th, w = I_th=0.15, w=0.05
Sigmoid threshold and transition width in Eq. (17), chosen for this study; they affect the reported |delta d| metric.

axioms (5)

domain assumption The background-decomposed PSFD equation with operator L and source b_bg is a valid discretization of Maxwell's equations for EUV mask scattering.
The entire loss and reference rest on this equation, taken from Ref. [24] without an independent benchmark here.
domain assumption The TMM-based reflection operator R_ml exactly represents the planar multilayer response for arbitrary scattered fields at z=z_ml.
Used in Eqs. (6)-(7); assumes planar stack, up/downgoing and TE/TM decomposition, and no mode coupling in the background.
domain assumption Gaussian smoothing of the binary mask yields a material distribution whose scattered field approximates the physical mask with acceptable accuracy.
Applied to both training and reference; smoothing width and its effect on lithographic accuracy are not quantified.
domain assumption The factorized FNO of finite width/depth can represent the solution operator of the vectorial Helmholtz equation well enough for the surrogate to be accurate.
This is the central approximation; empirically supported only by the reported MAE on 100 masks.
standard math BiCGSTAB with the described right preconditioner converges for the complex non-Hermitian PSFD system.
Standard Krylov method assumption; convergence is empirical.

pith-pipeline@v1.3.0-alltime-deepseek · 12313 in / 16454 out tokens · 167440 ms · 2026-08-01T02:44:23.559739+00:00 · methodology

0 comments

read the original abstract

We present a physics-informed neural operator (PINO) trained with pseudo-spectral frequency-domain (PSFD) equations for electromagnetic (EM) scattering problems in EUV lithography. The Fourier neural operator is factorized into a two-dimensional lateral ($xy$) branch and a one-dimensional axial ($z$) branch and is trained self-consistently with background decomposition.Thus, the full-vector coupling between the mask and the multilayer response is retained without invoking a finite-order Born approximation. In this way, the computational domain size is significantly reduced, thereby lowering the computational cost. The PINO is trained on approximately 16,000 mask designs from the LithoBench library sampled randomly at each training iteration without using precomputed EM field solutions. The PINO surrogate model yields predictions with a mean absolute error of about $7 \times 10^{-3}$ for the scattered intensity of held-out mask patterns relative to the reference PSFD solution. Combined with spectral damping, the PINO warm-start initialization accelerates the background-decomposed PSFD solver on finer discretizations.

Figures

Figures reproduced from arXiv: 2607.25330 by Doyun Kim, Werner Gillijns.

**Figure 1.** Figure 1: (a) Scattered intensity |𝐸 scat 𝑦 | 2 computed by the background-decomposed PSFD solver with the computational domain covering only the 50 nm Ta absorber region (dashed outline), (b) Vertical profile of |𝐸 scat 𝑦 | 2 along the 𝑧-axis at the center of the absorber region for both solvers at ∆𝑥 = 1 nm, (c, d) TE reflectance |𝑟TE| 2 and reflection phase arg(𝑟TE) of the bare multilayer computed by the standard… view at source ↗

**Figure 2.** Figure 2: (a) Regularized input field encoding the material distribution of a representative test [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Regularized input field of a representative held-out mask gemetry, displayed as two [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison of the scattered intensity [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: (a) Relative residual as a function of iterations for the three solver configurations. (b) [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Convergence trends of the background-decomposed PSFD solver under three configu [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

**Figure 7.** Figure 7: Spatial maps of the wafer-equivalent near-field width error [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 1 canonical work pages · 1 internal anchor

[1]

A.Taflove, AdvancesinComputationalElectrodynamics: TheFinite-difference Time-domainMethod,ArtechHouse(1998)

1998
[2]

W.C.Chew, WavesandFieldsinInhomogeneousMedia ,Wiley-IEEEPress,NewYork (1995)

1995
[3]

14(1966)302

K.Yee, Numericalsolutionofinitialboundaryvalueproblemsinvolvingmaxwell’sequations inisotropicmedia ,IEEETrans.AntennasPropagat. 14(1966)302

1966
[4]

M.G.Moharam,E.B.Grann,D.A.PommetandT.K.Gaylord, Formulationforstableand efficientimplementationoftherigorouscoupled-waveanalysisofbinarygratings ,Journalof theOpticalSocietyofAmericaA 12(1995)1068

1995
[5]

H.H.Hopkins, Onthediffractiontheoryofopticalimages ,ProceedingsoftheRoyalSocietyof London.SeriesA.MathematicalandPhysicalSciences 217(1953)408

1953
[6]

A.K.Wong, ResolutionEnhancementTechniquesinOpticalLithography ,SPIEPress, Bellingham,WA(2001)

2001
[7]

T.V.Pistor,K.AdamandA.Neureuther, Rigoroussimulationofmaskcornereffectsin extremeultravioletlithography ,J.Vac.Sci.Technol.B 16(1998)3449

1998
[8]

A.ErdmannandP.Evanschitzky, Modelingofmaskeffectsinopticalandeuvlithography ,in PhotomaskandNext-GenerationLithographyMaskTechnologyXIII ,vol.6283,p.62831B, SPIE,2006

2006
[9]

A.Erdmann, Mask3deffectsineuvlithography ,in EmergingLithographicTechnologiesIX , vol.5751,pp.553–564,SPIE,2005

2005
[10]

T.V.Pistor,Electromagneticsimulationofextremeultravioletlithographymasks ,in Emerging LithographicTechnologiesVI ,vol.4688,pp.1031–1041,SPIE,2002

2002
[11]

Three-dimensionalmaskmodelforphotolithography simulation

P.Liu,Y.Cao,L.ChenandJ.Ye,“Three-dimensionalmaskmodelforphotolithography simulation.”USPatent7703069B1,2010. 15

2010
[12]

K.AdamandA.R.Neureuther, Domaindecompositionmethodforrapidcalculationof photomasknearfields ,JournalofVacuumScience&TechnologyB:Microelectronicsand NanometerStructuresProcessing,Measurement,andPhenomena 20(2002)2585

2002
[13]

T.Watanabe,H.Suzuki,M.Takahashi,T.MatsunawaandS.Nojima, Fastmasknear-field calculationmethodusingdeeplearning ,JournalofMicro/Nanolithography,MEMS,and MOEMS18(2019)044001

2019
[14]

N.JiaandE.Y.Lam, Fastaerialimagecalculationforextremeultravioletlithographyunder partlycoherentillumination ,JournalofVacuumScience&TechnologyB 32(2014)021601

2014
[15]

M.Raissi,P.PerdikarisandG.E.Karniadakis, Physics-informedneuralnetworks: Adeep learningframeworkforsolvingforwardandinverseproblemsinvolvingnonlinearpartial differentialequations,JournalofComputationalPhysics 378(2019)686

2019
[16]

10.48550/arXiv.2111.03794

Z.Li,H.Zheng,N.Kovachki,D.Jin,H.Chen,B.Liuetal., Physics-InformedNeural OperatorforLearningPartialDifferentialEquations ,2023. 10.48550/arXiv.2111.03794

Pith/arXiv arXiv doi:10.48550/arxiv.2111.03794 2023
[17]

A.Erdmann,C.SchwemmerandP.Evanschitzky, Physics-informedneuralnetworksfor computationallithographymodelingandoptimization ,JournalofMicro/Nanopatterning, Materials,andMetrology 23(2024)011005

2024
[18]

V.Medvedev,A.ErdmannandA.Rosskopf, 3Dmasksimulationandlithographicimaging usingphysics-informedneuralnetworks ,in OpticalandEUVNanolithographyXXXVII , SPIE,2024,DOI

2024
[19]

V.A.Es’kinandE.V.Ivanov, Physics-informedneuralnetworksandneuraloperatorsfora studyofEUVelectromagneticwavediffractionfromalithographymask ,in 2025Dayson Diffraction(DD),2025,DOI

2025
[20]

J.Peurifoy,Y.Shen,L.Jing,Y.Yang,F.Cano-Garcia,M.Tegmarketal., Nanophotonic particlesimulationandinversedesignusingartificialneuralnetworks ,ScienceAdvances 4 (2018)eaar4206

2018
[21]

Y.Khoo,J.LuandL.Ying, Solvingparametricscatteringproblemsvianeuraloperators , JournalofComputationalPhysics 443(2021)110542

2021
[22]

P.EvanschitzkyandA.Erdmann, ThreedimensionalEUVsimulations: anewmasknear fieldandimagingsimulationsystem ,in EmergingLithographicTechnologiesIX ,SPIE,2005, DOI

2005
[23]

J.Zhang,Y.LiandL.Wang, Fastrigorousmaskmodelforextremeultravioletlithography , AppliedOptics (2020)

2020
[24]

Pseudo-spectral frequency-domain method with background field decomposition and Green's function preconditioner for electromagnetic scattering problem in EUV lithography

S.Lee,W.GillijnsandD.Kim, Pseudo-spectralfrequency-domainmethodwithbackground fielddecompositionandGreen’sfunctionpreconditionerforelectromagneticscattering probleminEUVlithography ,2026. 10.48550/arXiv.2606.25541

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.25541 2026
[25]

A.Tran,A.Mathews,L.XieandC.S.Ong, Factorizedfourierneuraloperators ,in InternationalConferenceonLearningRepresentations ,2023[ 2111.13802]. 16

Pith/arXiv arXiv 2023
[26]

10.48550/arXiv.2010.08895

Z.Li,N.Kovachki,K.Azizzadenesheli,B.Liu,K.Bhattacharya,A.Stuartetal., Fourier NeuralOperatorforParametricPartialDifferentialEquations ,2021. 10.48550/arXiv.2010.08895

Pith/arXiv arXiv doi:10.48550/arxiv.2010.08895 2021
[27]

10.5555/3648699.3648788

N.Kovachki,Z.Li,B.Liu,K.Azizzadenesheli,K.Bhattacharya,A.Stuartetal., Neural Operator: LearningMapsBetweenFunctionSpaces ,2024. 10.5555/3648699.3648788

arXiv 2024
[28]

I.A.Makhotkin,M.Wu,V.Soltwisch,F.ScholzeandV.Philipsen, Refinedextreme ultravioletmaskstackmodel ,J.Opt.Soc.Am.A 38(2021)498

2021
[29]

B.Fornberg, APracticalGuidetoPseudospectralMethods ,CambridgeUniversityPress (1998)

1998
[30]

Q.H.Liu, ThePSTDalgorithm: Atime-domainmethodrequiringonlytwocellsper wavelength,MicrowaveandOpticalTechnologyLetters 15(1997)158

1997
[31]

A.Farjadpour,D.Roundy,A.Rodriguez,M.Ibanescu,P.Bermel,J.D.Joannopoulosetal., Improvingaccuracybysubpixelsmoothinginthefinite-differencetimedomain ,OpticsLetters 31(2006)2972

2006
[32]

10.48550/arXiv.2309.00583

Z.Li,N.Kovachki,C.Choy,B.Li,J.Kossaifi,S.P.Ottaetal., Geometry-InformedNeural OperatorforLarge-Scale3DPDEs ,2023. 10.48550/arXiv.2309.00583

Pith/arXiv arXiv doi:10.48550/arxiv.2309.00583 2023
[33]

Y.Saad, IterativeMethodsforSparseLinearSystems ,SocietyforIndustrialandApplied Mathematics,Philadelphia,PA,2nded.(2003),10.1137/1.9780898718003

doi:10.1137/1.9780898718003 2003
[34]

13(1992)631

H.A.vanderVorst, Bi-CGSTAB:AFastandSmoothlyConvergingVariantofBi-CGforthe SolutionofNonsymmetricLinearSystems ,SIAMJ.Sci.Stat.Comput. 13(1992)631

1992
[35]

M.Benzi, Preconditioningtechniquesforlargelinearsystems: Asurvey ,Journalof ComputationalPhysics 182(2002)418. 17

2002

[1] [1]

A.Taflove, AdvancesinComputationalElectrodynamics: TheFinite-difference Time-domainMethod,ArtechHouse(1998)

1998

[2] [2]

W.C.Chew, WavesandFieldsinInhomogeneousMedia ,Wiley-IEEEPress,NewYork (1995)

1995

[3] [3]

14(1966)302

K.Yee, Numericalsolutionofinitialboundaryvalueproblemsinvolvingmaxwell’sequations inisotropicmedia ,IEEETrans.AntennasPropagat. 14(1966)302

1966

[4] [4]

M.G.Moharam,E.B.Grann,D.A.PommetandT.K.Gaylord, Formulationforstableand efficientimplementationoftherigorouscoupled-waveanalysisofbinarygratings ,Journalof theOpticalSocietyofAmericaA 12(1995)1068

1995

[5] [5]

H.H.Hopkins, Onthediffractiontheoryofopticalimages ,ProceedingsoftheRoyalSocietyof London.SeriesA.MathematicalandPhysicalSciences 217(1953)408

1953

[6] [6]

A.K.Wong, ResolutionEnhancementTechniquesinOpticalLithography ,SPIEPress, Bellingham,WA(2001)

2001

[7] [7]

T.V.Pistor,K.AdamandA.Neureuther, Rigoroussimulationofmaskcornereffectsin extremeultravioletlithography ,J.Vac.Sci.Technol.B 16(1998)3449

1998

[8] [8]

A.ErdmannandP.Evanschitzky, Modelingofmaskeffectsinopticalandeuvlithography ,in PhotomaskandNext-GenerationLithographyMaskTechnologyXIII ,vol.6283,p.62831B, SPIE,2006

2006

[9] [9]

A.Erdmann, Mask3deffectsineuvlithography ,in EmergingLithographicTechnologiesIX , vol.5751,pp.553–564,SPIE,2005

2005

[10] [10]

T.V.Pistor,Electromagneticsimulationofextremeultravioletlithographymasks ,in Emerging LithographicTechnologiesVI ,vol.4688,pp.1031–1041,SPIE,2002

2002

[11] [11]

Three-dimensionalmaskmodelforphotolithography simulation

P.Liu,Y.Cao,L.ChenandJ.Ye,“Three-dimensionalmaskmodelforphotolithography simulation.”USPatent7703069B1,2010. 15

2010

[12] [12]

K.AdamandA.R.Neureuther, Domaindecompositionmethodforrapidcalculationof photomasknearfields ,JournalofVacuumScience&TechnologyB:Microelectronicsand NanometerStructuresProcessing,Measurement,andPhenomena 20(2002)2585

2002

[13] [13]

T.Watanabe,H.Suzuki,M.Takahashi,T.MatsunawaandS.Nojima, Fastmasknear-field calculationmethodusingdeeplearning ,JournalofMicro/Nanolithography,MEMS,and MOEMS18(2019)044001

2019

[14] [14]

N.JiaandE.Y.Lam, Fastaerialimagecalculationforextremeultravioletlithographyunder partlycoherentillumination ,JournalofVacuumScience&TechnologyB 32(2014)021601

2014

[15] [15]

M.Raissi,P.PerdikarisandG.E.Karniadakis, Physics-informedneuralnetworks: Adeep learningframeworkforsolvingforwardandinverseproblemsinvolvingnonlinearpartial differentialequations,JournalofComputationalPhysics 378(2019)686

2019

[16] [16]

10.48550/arXiv.2111.03794

Z.Li,H.Zheng,N.Kovachki,D.Jin,H.Chen,B.Liuetal., Physics-InformedNeural OperatorforLearningPartialDifferentialEquations ,2023. 10.48550/arXiv.2111.03794

Pith/arXiv arXiv doi:10.48550/arxiv.2111.03794 2023

[17] [17]

A.Erdmann,C.SchwemmerandP.Evanschitzky, Physics-informedneuralnetworksfor computationallithographymodelingandoptimization ,JournalofMicro/Nanopatterning, Materials,andMetrology 23(2024)011005

2024

[18] [18]

V.Medvedev,A.ErdmannandA.Rosskopf, 3Dmasksimulationandlithographicimaging usingphysics-informedneuralnetworks ,in OpticalandEUVNanolithographyXXXVII , SPIE,2024,DOI

2024

[19] [19]

V.A.Es’kinandE.V.Ivanov, Physics-informedneuralnetworksandneuraloperatorsfora studyofEUVelectromagneticwavediffractionfromalithographymask ,in 2025Dayson Diffraction(DD),2025,DOI

2025

[20] [20]

J.Peurifoy,Y.Shen,L.Jing,Y.Yang,F.Cano-Garcia,M.Tegmarketal., Nanophotonic particlesimulationandinversedesignusingartificialneuralnetworks ,ScienceAdvances 4 (2018)eaar4206

2018

[21] [21]

Y.Khoo,J.LuandL.Ying, Solvingparametricscatteringproblemsvianeuraloperators , JournalofComputationalPhysics 443(2021)110542

2021

[22] [22]

P.EvanschitzkyandA.Erdmann, ThreedimensionalEUVsimulations: anewmasknear fieldandimagingsimulationsystem ,in EmergingLithographicTechnologiesIX ,SPIE,2005, DOI

2005

[23] [23]

J.Zhang,Y.LiandL.Wang, Fastrigorousmaskmodelforextremeultravioletlithography , AppliedOptics (2020)

2020

[24] [24]

Pseudo-spectral frequency-domain method with background field decomposition and Green's function preconditioner for electromagnetic scattering problem in EUV lithography

S.Lee,W.GillijnsandD.Kim, Pseudo-spectralfrequency-domainmethodwithbackground fielddecompositionandGreen’sfunctionpreconditionerforelectromagneticscattering probleminEUVlithography ,2026. 10.48550/arXiv.2606.25541

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.25541 2026

[25] [25]

A.Tran,A.Mathews,L.XieandC.S.Ong, Factorizedfourierneuraloperators ,in InternationalConferenceonLearningRepresentations ,2023[ 2111.13802]. 16

Pith/arXiv arXiv 2023

[26] [26]

10.48550/arXiv.2010.08895

Z.Li,N.Kovachki,K.Azizzadenesheli,B.Liu,K.Bhattacharya,A.Stuartetal., Fourier NeuralOperatorforParametricPartialDifferentialEquations ,2021. 10.48550/arXiv.2010.08895

Pith/arXiv arXiv doi:10.48550/arxiv.2010.08895 2021

[27] [27]

10.5555/3648699.3648788

N.Kovachki,Z.Li,B.Liu,K.Azizzadenesheli,K.Bhattacharya,A.Stuartetal., Neural Operator: LearningMapsBetweenFunctionSpaces ,2024. 10.5555/3648699.3648788

arXiv 2024

[28] [28]

I.A.Makhotkin,M.Wu,V.Soltwisch,F.ScholzeandV.Philipsen, Refinedextreme ultravioletmaskstackmodel ,J.Opt.Soc.Am.A 38(2021)498

2021

[29] [29]

B.Fornberg, APracticalGuidetoPseudospectralMethods ,CambridgeUniversityPress (1998)

1998

[30] [30]

Q.H.Liu, ThePSTDalgorithm: Atime-domainmethodrequiringonlytwocellsper wavelength,MicrowaveandOpticalTechnologyLetters 15(1997)158

1997

[31] [31]

A.Farjadpour,D.Roundy,A.Rodriguez,M.Ibanescu,P.Bermel,J.D.Joannopoulosetal., Improvingaccuracybysubpixelsmoothinginthefinite-differencetimedomain ,OpticsLetters 31(2006)2972

2006

[32] [32]

10.48550/arXiv.2309.00583

Z.Li,N.Kovachki,C.Choy,B.Li,J.Kossaifi,S.P.Ottaetal., Geometry-InformedNeural OperatorforLarge-Scale3DPDEs ,2023. 10.48550/arXiv.2309.00583

Pith/arXiv arXiv doi:10.48550/arxiv.2309.00583 2023

[33] [33]

Y.Saad, IterativeMethodsforSparseLinearSystems ,SocietyforIndustrialandApplied Mathematics,Philadelphia,PA,2nded.(2003),10.1137/1.9780898718003

doi:10.1137/1.9780898718003 2003

[34] [34]

13(1992)631

H.A.vanderVorst, Bi-CGSTAB:AFastandSmoothlyConvergingVariantofBi-CGforthe SolutionofNonsymmetricLinearSystems ,SIAMJ.Sci.Stat.Comput. 13(1992)631

1992

[35] [35]

M.Benzi, Preconditioningtechniquesforlargelinearsystems: Asurvey ,Journalof ComputationalPhysics 182(2002)418. 17

2002