Pith. sign in

REVIEW 3 major objections 7 minor 70 references

No single flow-surrogate architecture wins both a boundary-driven film and a self-sustained wake; how the model treats time decides the winner, and pointwise error ranks the wrong model in both regimes.

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 · grok-4.5

2026-07-30 16:20 UTC pith:RR3QX5A3

load-bearing objection Solid two-regime bake-off with a real practice punchline: time treatment flips the winner, and RMSE lies in both regimes—worth engaging, with the encoder confound already half-owned by the authors. the 3 major comments →

arxiv 2607.23667 v1 pith:RR3QX5A3 submitted 2026-07-26 math.NA cs.LGcs.NAphysics.flu-dyn

No Free Lunch in Flow Surrogates under Time-Varying Boundary Conditions: A Two-Regime Study

Georg Winkler , Martin Stoll This is my paper
classification math.NA cs.LGcs.NAphysics.flu-dyn MSC 65M6076D0568T07
keywords surrogate modelingreduced-order modelingoperator learningDeepONetKármán vortex streetchemical-mechanical planarisationtime-varying boundary conditionsfailure-mode metrics
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.

Engineers often treat a surrogate that works on a simplified flow as proof it will work on a richer one. This paper tests that habit on two startup-driven flows: a quasi-static three-dimensional slurry film from chemical-mechanical planarisation, and a two-dimensional Kármán vortex street that sheds on its own. Eight models that either map the full field or a latent state, and either predict the whole trajectory at once or step by step, are scored on one shared pipeline. No architecture wins both. A one-shot full-field map reconstructs the film’s process target—cumulative wall shear—to about three percent relative error, while latent autoregressive models keep nearly all of the wake’s shedding power that direct and one-shot models damp away. The deciding choice is the treatment of time: the self-sustained wake needs the phase memory of feedback, while the boundary-driven film rewards a direct map. Aggregate root-mean-square error inverts the physically relevant ranking in both regimes, so the paper scores field accuracy, structure, invented motion, amplitude, and timing instead. The practical message is that surrogate choice should follow the flow’s dynamical character, validation should resolve failure modes, and the speedup only pays off after enough queries to amortise the training simulations.

Core claim

Among eight surrogates compared on a shared pipeline, no single architecture wins both regimes. On the CMP film a one-shot full-field DeepONet reaches 3.2% relative error on cumulative wall shear stress; on the Kármán wake a latent autoregressive DeepONet retains about 96% of the shedding power that direct and one-shot models collapse to nearly zero. The axis that flips the winner is the treatment of time—autoregressive feedback for the self-sustained limit cycle, a direct map for the boundary-driven Stokes film—while representation only changes the margin. Pointwise RMSE ranks the wrong model in both regimes, so failure-mode-resolved metrics are required; neither the winning architecture no

What carries the argument

A two-axis design matrix of representation (full-field versus frozen latent encoder) against time treatment (one-shot map versus autoregressive rollout), evaluated by a fixed suite of five physical questions—field accuracy, spatial structure, invented motion, fluctuation amplitude, and event timing—rather than by aggregate RMSE alone.

Load-bearing premise

That contrasting these two flows is enough to pin the architecture flip on dynamical character alone, even though the regimes also differ in encoder, dimension, governing equations, and how the ramp family is built.

What would settle it

Hold the encoder, dimension, and ramp family fixed and swap only whether the target flow is boundary-driven or self-sustained; if the one-shot versus autoregressive ranking no longer flips with that dynamical character, the central claim fails.

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

If this is right

  • A CMP surrogate programme cannot treat success on a simplified Stokes film as evidence that the same architecture will hold once richer dynamics appear.
  • Self-sustained oscillatory flows should be assigned autoregressive latent models; boundary-driven quasi-static flows should be assigned one-shot direct maps.
  • Model selection by pointwise RMSE alone will deploy the damped wake predictor and a film model several times worse on the process target.
  • Surrogates pay off only as many-query instruments: break-even is set almost entirely by training-set size (about 70 queries for the film, about 647 for the wake), not by training compute.
  • Validation reports should include failure-mode-resolved scores for structure, hallucination, amplitude, and timing alongside any aggregate error.

Where Pith is reading between the lines

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

  • The same time-treatment split is likely to reappear in other manufacturing flows that mix a forced ramp with possible spontaneous oscillation, such as coating, filling, or stirred reactors.
  • Once a process model adds particle transport or free-surface dynamics on top of the film, the winning cell may flip mid-programme, so architecture should be re-chosen when the physics enrichment crosses into self-sustained behaviour.
  • Active learning or smaller training sets could move the break-even count down, but only if accuracy saturates early enough that the offline simulation bill shrinks without losing the physical metrics.
  • Unpinning of mirror limit-cycle branches is a silent deployment risk for any wake-like surrogate scored only by pointwise error.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The manuscript compares eight surrogate architectures — spanning direct vs latent representation and one-shot vs autoregressive time traversal — on two transient flows driven by piecewise-linear boundary-condition ramps: a quasi-static 3D Stokes slurry film motivated by chemical-mechanical planarisation (CMP), and the 2D Kármán vortex street behind a cylinder. The central finding is a no-free-lunch result: a one-shot full-field DeepONet wins the film on the process target (cumulative wall shear stress, cwL2≈0.032) while latent autoregressive DeepONets (S4D/Mamba/GRU) win the wake (retaining up to 96% of shedding power where direct and one-shot models damp it to ~0). The authors attribute the reversal to the treatment of time: the self-sustained limit cycle needs autoregressive phase memory, the boundary-driven film rewards a direct map. A second contribution is a five-aspect failure-mode metric suite (field, structure, invented motion, amplitude, timing), with the demonstration that pointwise RMSE inverts the physically relevant ranking in both regimes. A third is a break-even cost analysis showing the surrogate pays off only beyond ~70 (film) and ~647 (wake) queries. Models share splits, frozen encoders within regime, fixed epoch budgets, three seeds, and metrics fixed before ranking.

Significance. If the results hold, this is a useful and unusually careful empirical contribution to operator learning for transient PDEs. The concrete strengths: (i) a controlled within-regime ablation isolating time treatment (frozen shared encoder, same branch family, one-shot vs autoregressive), which supports the central causal reading far better than the cross-regime contrast alone; (ii) a failure-mode-resolved metric suite fixed before model ranking, with a demonstrated, quantified inversion of model ranking under pointwise RMSE in both regimes (Table 2: nRMSE3 ranks GRU-DON ahead of the field DeepONet while cwL2 separates them 4× in the opposite direction; Table 3: the damped model wins RMSE at P=0.004 vs 0.96); (iii) the mirror-branch diagnostic (§7), which cleanly explains the S4D RMSE deficit as an unpinned symmetry choice (1.10→0.68 under mirror-rescoring) — a falsifiable, mechanistic check; and (iv) honest cost accounting with break-even query counts Q*≈70/647 (Eq. 27), directly answering the McGreivy–Hakim reporting critique. The CMP motivation is real and the "validation does not transfer across regimes" caution is practically relevant to surrogate programmes. The claims are approp

major comments (3)
  1. [§4.2, Table 2] §4.2 / Table 2 / Fig. 7: The film-side headline gap (direct DeepONet cwL2=0.032 vs best latent-AR 0.144) is interpretable as a time-treatment effect only if the POD k=8 encoder does not itself set a cwL2 floor well above 0.032. The truncation is chosen by an energy criterion (98.8%, Fig. 5), but the target S(x)=∫|τw|dt (Eq. 3) divides the velocity difference by h=40 µm and integrates in time, so it can be dominated by the discarded 1.2% if that energy sits in near-wall, high-gradient structure. The latent models cannot beat their encoder's reconstruction floor regardless of dynamics quality, so part of the >4× gap may measure representation, not the direct map. There is a cheap, decisive control that requires no retraining: project the reference test fields through the frozen POD k=8 basis and report cwL2, corr, and ΔWIWNU of the encode–decode reconstruction. If the floor is ≪0.1, the ca
  2. [§7, Table 1] §7, Table 1: The claim that 'the treatment of time decides who wins and the representation influences the margin' is, on the wake, established only within the latent family: the direct×autoregressive cell is empty because neither the autoregressive field DeepONet nor the FNO variant could be trained stably (§8). A direct-AR model that held the shedding would break the clean attribution (it would suggest representation/time interact rather than separate). The failed-training disclosure is commendable, but the conclusion should be scoped accordingly: on the wake, autoregressive feedback is shown necessary for latent models and sufficient given the frozen conv-AE, while for full-field models only the one-shot failure is demonstrated. Please qualify the abstract and §7 mechanism paragraph (e.g., 'no representation rescues the one-shot map' is supported; the converse necessity for direct mode
  3. [§5] §5.1–5.2: Several suite instruments carry free thresholds — the hallucination velocity threshold τ=0.1×peak (Eq. 16), the onset detector's 0.1-of-developed-fluctuation bar, and the DTW alignment window in RMSEw (Eq. 15) — and the text asserts the detector parameters 'affect the model comparison only marginally' without evidence. The central RMSE-inversion claim rests mainly on P (Eq. 21) and cwL2 (Eq. 20), which are threshold-free, so the risk is contained; still, since the suite is itself a contribution and is recommended for adoption by CMP surrogate programmes, a small sensitivity table (τ and onset threshold varied over a plausible range, e.g. 0.05–0.2, reporting H and Δt∗ rankings for the leading models) would substantiate the invariance claim rather than assert it.
minor comments (7)
  1. [§3.1] §3.1: The difference-quotient shear τw=μ(u_top−u_bot)/h is said to agree with the resolved wafer-side wall shear 'to within a few percent across the pad land' — please give the actual number or a figure reference, since Eq. (20) scores surrogates against this convention.
  2. [Table 2] §5.3/Table 2: Film metrics are means over only 15 test cases per split, and Fig. 8 shows cwL2 ranging from 0.011 to 0.139 across cases. Please report the per-case distribution (median/IQR) for the headline cwL2 numbers, not only the seed spread, so the reader can separate case variance from seed variance.
  3. [§4.2] Fig. 5 vs §4.2: The conv-AE latent dimension 32 for KVS coincides almost exactly with the k=31 at which POD reaches 99%; one sentence on whether this was chosen by that coincidence or independently would help.
  4. [§2] §2: The claim that a flow driven by a boundary condition varying during the transient 'remains to be addressed' should be tempered slightly — e.g., [43] and time-dependent-input DeepONet variants [59] are adjacent; the novelty is specifically the ramp-BC startup setting with the two-regime contrast.
  5. [Throughout] Typos/typesetting: abstract 'pay offfrom' (missing space); inconsistent spacing around numbers ('0 .77 h', '103 to 104' for 10^3–10^4 in abstract); 'Reh' renders with subscript spacing issues; 'Sec.' vs 'Section' usage varies.
  6. [Table 3] Table 3: for DMDc and S4D one-shot, Δt∗ is reported as '—' with a footnote ('sheds on fewer than half of the test cases'); please state in the caption what fraction of cases those models do shed, since 'undefined onset' is itself informative about the damping failure mode.
  7. [§6.3] §6.3/Table 4: The CPU-vs-GPU speedup S is honestly flagged as a deployment comparison; please also state the GPU inference batch size used for tinfer, since per-case latency at batch 1 vs full batch can differ by an order of magnitude and Q∗ depends on it weakly through Eq. (27).

Circularity Check

0 steps flagged

Empirical architecture bake-off against external FEM references; no prediction reduces to its inputs by construction.

full rationale

The paper’s central claims are comparative experimental outcomes on two FEM-generated datasets, not first-principles derivations. Training losses are MSE on field or latent states only; evaluation metrics (cwL2, shedding power P, hallucination H, onset latency, etc.) never enter the loss (§4.3), so process-target and dynamics wins are not optimized in by construction. Metrics were fixed from early failure-mode inspection before the model comparison (§5). Reference solutions come from independent DOLFINx FEM runs; break-even Q* follows from measured wall-clocks and N_train via the standard amortisation formula. The sole self-citation ([2], authors’ CMP ML review) supplies a schematic and survey context and is not load-bearing for the no-free-lunch ranking. Encoder–regime confounding (§8) is a validity/causal-identification concern, not circularity. No fitted parameter is renamed a prediction, no uniqueness theorem is imported from the authors, and no ansatz is smuggled in via self-citation.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 2 invented entities

The claim is empirical, not axiomatic physics. It rests on standard continuum models, standard ML building blocks, and several modeling choices that define the two ‘regimes’ and the fairness of the bake-off. Free parameters are mostly architectural/training knobs and metric thresholds that could move margins; invented entities are evaluation constructs, not new physical objects.

free parameters (7)
  • POD truncation k=8 (CMP latent) = k=8
    Chosen by energy truncation (~98.8% EVR); changes latent capacity and rollout stability on the film.
  • Conv-AE latent dimension 32 (KVS) = 32
    Nonlinear compressor size for the wake; set above POD 99% mode count (31), not cross-validated against all dynamics models.
  • Hallucination velocity threshold τ=0.1×peak steady |u| = 0.1
    Hand threshold defining pre-onset invented motion H; rescaling changes absolute H though relative model order may be stable.
  • Onset detector threshold (0.1 of developed stripe fluctuation) = 0.1
    Defines t* and Δt*; paper says absolute times shift but comparisons only marginally—still a free detector knob.
  • Fixed epoch budgets and home-regime hyperparameter transfer
    Budgets set heuristically until validation plateau; champions tuned on home regime and transferred (§4.3), which can shift margins though authors argue it cannot create the flip.
  • Variance penalty weight in AR rollout training
    Extra loss term to stop collapse to temporal mean; strength is a training choice affecting held shedding power.
  • Ramp family shape (levels, α, knot Dirichlet partitions) = CMP: 4 levels step 0.25 α=1; KVS: 5 levels step 0.2 α=2
    Defines the trained input distribution; generalization is only claimed inside this piecewise-linear family (§3, §8).
axioms (6)
  • domain assumption Incompressible Navier–Stokes / unsteady Stokes with the stated BCs and FEM discretizations are adequate reference physics for the two regimes.
    §3: film reduced to unsteady Stokes by Re_h(h/L)≪1; wake uses full NS on Schäfer–Turek geometry. Claim compares surrogates to these references, not to plant data.
  • domain assumption Cumulative wall shear stress (time-integrated |τ_w|) is the process-relevant film target linking to material removal.
    §3.1 via Preston/shear-force laws; hydrodynamic part only, fixed geometry—accepted engineering proxy, not full CMP.
  • ad hoc to paper Shedding power P, phase-aligned RMSEw, H/α_rms, and onset latency are the right physical questions for ranking wake surrogates.
    §5 constructs the suite from early visual failure modes; standard in spirit but specific definitions and thresholds are paper-local.
  • ad hoc to paper Within-regime frozen encoder plus shared pipeline makes architecture differences attributable to representation/time axes rather than co-adapted compression.
    §4.2–4.3; cross-regime encoder difference is intentionally confounded and limits causal cross-regime claims (§8).
  • domain assumption Break-even Q* ≈ N_train when t_infer ≪ t_FEM is the right economic comparison of surrogate vs solver.
    §5.4 following McGreivy & Hakim reporting advice; search/development cost excluded by design.
  • standard math Standard approximation/properties of DeepONet, S4D, Mamba, GRU, DMDc, and POD/AE as function/sequence approximators.
    Used as given from cited ML literature; no new approximation theorems proved.
invented entities (2)
  • Five-aspect failure-mode metric suite (field, structure, invented motion, amplitude, timing) with regime-specific instruments no independent evidence
    purpose: Replace aggregate RMSE that inverts physical rankings; score damping, hallucination, onset lag, cum-WSS, etc.
    Operational definitions (cwL2, D, P, H, α_rms, Δt*, ESt, ΔWIWNU, RMSEw) are introduced here as the evaluation standard for the bake-off.
  • Two-regime CMP-motivated no-free-lunch framing (film vs wake under shared ramp-BC construction) no independent evidence
    purpose: Test whether simple-regime validation transfers to richer dynamics relevant to future CMP models.
    Pairing and ramp construction are study design, not a new physical law; evidence is internal to the experiments.

pith-pipeline@v1.2.0-grok45-kimik3 · 31665 in / 4465 out tokens · 86881 ms · 2026-07-30T16:20:29.599024+00:00 · methodology

0 comments
read the original abstract

A flow surrogate validated on a simple regime is often taken as evidence that the approach will carry to a richer one. We test this assumption on two transient flows under time-varying boundary conditions emulating the process startup: the three-dimensional slurry film in chemical-mechanical planarisation (CMP), a core semiconductor-manufacturing process, and the two-dimensional Karman vortex street (KVS) behind a cylinder. Eight surrogate models are compared on one shared evaluation pipeline, differing in whether they learn the full field or a latent representation, and whether they predict trajectories in one shot or step by step. No single architecture wins both regimes. On the film, a one-shot full-field model reconstructs the process-relevant cumulative wall shear stress to 3.2% relative error. On the wake, a latent autoregressive DeepONet retains 96% of the shedding power that direct and one-shot models damp to almost zero. The deciding axis is the treatment of time. The self-sustained wake requires the phase memory that autoregressive feedback provides, while the boundary-driven film rewards a direct map. Pointwise RMSE picks the wrong model in both regimes, so the evaluation scores five physical questions instead, the field, its structure, invented motion, amplitude, and timing. The trained surrogates answer queries $10^3$ to $10^4$ times faster than the finite-element solver, but the offline cost of the training simulations means they pay off from the first query beyond the training set for CMP and the third for the KVS. The choice of surrogate should follow the dynamical character of the target flow, and its validation should use failure-mode-resolved metrics, since neither the winning architecture nor its validation transfers.

Figures

Figures reproduced from arXiv: 2607.23667 by Georg Winkler, Martin Stoll.

Figure 1
Figure 1. Figure 1: Schematic illustration of the CMP process: a rotating wafer (held by [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Temporal evolution of the ground-truth finite element method solution under the boundary-condition ramp. (a) CMP: the cumulative wall shear stress [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The piecewise-linear boundary-condition ramp (representative KVS schedules). (a) Six example ramps [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: KVS setup, following the Schäfer–Turek benchmark [ [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cumulative POD energy of the two regimes (train split). CMP concen [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The four KVS failure modes that an aggregate RMSE conflates. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: CMP model comparison on the interpolation split, on the process [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: CMP cum-WSS field of the direct operator (field DeepONet) against truth on its best and worst interpolation case by [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: KVS model comparison on the interpolation split. (a) Shedding power [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗

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

70 extracted references · 10 canonical work pages

  1. [1]

    P. B. Zantye, A. Kumar, A. K. Sikder, Chemical mechani- cal planarization for microelectronics applications, Materi- als Science and Engineering: R: Reports 45 (3-6) (2004) 89–220.doi:10.1016/j.mser.2004.06.002

    work page doi:10.1016/j.mser.2004.06.002 2004

  2. [2]

    Winkler, T

    G. Winkler, T. Rothe, M. A. Sayyed, L. Jäckel, J. Langer, H. Kuhn, M. Stoll, Machine learning in chemical– mechanical planarization: A comprehensive review of trends, applications, and challenges, Advanced Engineer- ing Informatics 68 (2025) 103663. doi:10.1016/j.aei. 2025.103663

    arXiv 2025

  3. [3]

    Krishnan, J

    M. Krishnan, J. W. Nalaskowski, L. M. Cook, Chemical mechanical planarization: Slurry chemistry, materials, and mechanisms, Chemical Reviews 110 (1) (2010) 178–204. doi:10.1021/cr900170z

    work page doi:10.1021/cr900170z 2010

  4. [4]

    S. Kim, J. Jang, C. O. Kim, A run-to-run controller for a chemical mechanical planarization process using least squares generative adversarial networks, Journal of In- telligent Manufacturing 32 (8) (2021) 2267–2280. doi: 10.1007/s10845-020-01639-1

    work page doi:10.1007/s10845-020-01639-1 2021

  5. [5]

    Cheng, A

    C. Cheng, A. Sa-Ngasoongsong, O. Beyca, T. Le, H. Yang, Z. Kong, S. T. Bukkapatnam, Time series forecasting for nonlinear and non-stationary processes: a review and com- parative study, IIE Transactions 47 (10) (2015) 1053–1071. doi:10.1080/0740817X.2014.999180. 15

    arXiv 2015

  6. [6]

    T. Pan, J. Yang, W. Wu, Z. Dong, B. O. Ayinde, Arti- ficial intelligent matching for scratches of semiconduc- tor wafers based on a k-NN algorithm, Surface Topog- raphy: Metrology and Properties 7 (2) (2019) 025012. doi:10.1088/2051-672X/ab1ca3

    work page doi:10.1088/2051-672x/ab1ca3 2019

  7. [7]

    M. A. Jebri, G. Graton, E. M. El Adel, M. Ouladsine, J. Pinaton, Virtual metrology on chemical mechanical planarization process based on just-in-time learning, in: 2016 5th International Conference on Systems and Control (ICSC), 2016, pp. 169–174. doi:10.1109/ICoSC.2016. 7507082

    work page doi:10.1109/icosc.2016 2016

  8. [8]

    Q. Xu, L. Chen, H. Cao, J. Liu, A neural network-based approach to material removal rate prediction for copper chemical mechanical planarization, ECS Journal of Solid State Science and Technology 10 (5) (2021) 054003. doi: 10.1149/2162-8777/abfc20

    work page doi:10.1149/2162-8777/abfc20 2021

  9. [9]

    J. Deng, Q. Zhang, J. Lu, Q. Yan, J. Pan, R. Chen, Predic- tion of the surface roughness and material removal rate in chemical mechanical polishing of single-crystal SiC via a back-propagation neural network, Precision Engineering 72 (2021) 102–110. doi:10.1016/j.precisioneng. 2021.04.012

    work page doi:10.1016/j.precisioneng 2021

  10. [10]

    J. Cai, C. Yan, Y . Ma, B. Yu, D. Zhou, X. Zeng, NeurFill: Migrating full-chip CMP simulators to neural networks for model-based dummy filling synthesis, in: 2021 58th ACM/IEEE Design Automation Conference (DAC), 2021, pp. 187–192.doi:10.1109/DAC18074.2021.9586325

    arXiv 2021

  11. [11]

    J. H. Ferziger, M. Peri´c, R. L. Street, Computational Meth- ods for Fluid Dynamics, 4th Edition, Springer, Cham, 2020. doi:10.1007/978-3-319-99693-6

    doi:10.1007/978-3-319-99693-6 2020

  12. [12]

    Berkooz, P

    G. Berkooz, P. Holmes, J. L. Lumley, The proper or- thogonal decomposition in the analysis of turbulent flows, Annual Review of Fluid Mechanics 25 (1993) 539–575. doi:10.1146/annurev.fl.25.010193.002543

    arXiv 1993

  13. [13]

    Benner, S

    P. Benner, S. Gugercin, K. Willcox, A survey of projection- based model reduction methods for parametric dynamical systems, SIAM Review 57 (4) (2015) 483–531. doi:10. 1137/130932715

    2015

  14. [14]

    K. Lee, K. T. Carlberg, Model reduction of dynamical sys- tems on nonlinear manifolds using deep convolutional au- toencoders, Journal of Computational Physics 404 (2020) 108973.doi:10.1016/j.jcp.2019.108973

    arXiv 2020

  15. [15]

    J. H. Seidman, G. Kissas, G. J. Pappas, P. Perdikaris, Vari- ational autoencoding neural operators, in: International Conference on Machine Learning (ICML), V ol. 202 of PMLR, 2023, pp. 30491–30522, arXiv:2302.10351

    Pith/arXiv arXiv 2023

  16. [16]

    Farenga, S

    N. Farenga, S. Fresca, S. Brivio, A. Manzoni, On latent dynamics learning in nonlinear reduced order modeling, Neural Networks 185 (2025) 107146. doi:10.1016/j. neunet.2025.107146

    arXiv 2025

  17. [17]

    Oommen, K

    V . Oommen, K. Shukla, S. Goswami, R. Dingreville, G. E. Karniadakis, Learning two-phase microstructure evolution using neural operators and autoencoder architectures, npj Computational Materials 8 (2022) 190. doi:10.1038/ s41524-022-00876-7

    2022

  18. [18]

    Serrano, T

    L. Serrano, T. X. Wang, E. Le Naour, J.-N. Vittaut, P. Galli- nari, AROMA: Preserving spatial structure for latent PDE modeling with local neural fields, in: Advances in Neural Information Processing Systems (NeurIPS), V ol. 37, 2024, pp. 13489–13521, arXiv:2406.02176

    Pith/arXiv arXiv 2024

  19. [19]

    T. Chen, H. Chen, Universal approximation to nonlin- ear operators by neural networks with arbitrary activation functions and its application to dynamical systems, IEEE Transactions on Neural Networks 6 (4) (1995) 911–917. doi:10.1109/72.392253

    doi:10.1109/72.392253 1995

  20. [20]

    N. B. Kovachki, S. Lanthaler, A. M. Stuart, Operator learn- ing: Algorithms and analysis, in: S. Mishra, A. Townsend (Eds.), Numerical Analysis Meets Machine Learning, V ol. 25 of Handbook of Numerical Analysis, Elsevier, 2024, pp. 419–467. doi:10.1016/bs.hna.2024.05. 009

    doi:10.1016/bs.hna.2024.05 2024

  21. [21]

    Bhattacharya, B

    K. Bhattacharya, B. Hosseini, N. B. Kovachki, A. M. Stu- art, Model reduction and neural networks for parametric PDEs, SMAI Journal of Computational Mathematics 7 (2021) 121–157.doi:10.5802/smai-jcm.74

    doi:10.5802/smai-jcm.74 2021

  22. [22]

    L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators, Na- ture Machine Intelligence 3 (3) (2021) 218–229. doi: 10.1038/s42256-021-00302-5

    doi:10.1038/s42256-021-00302-5 2021

  23. [23]

    Lanthaler, S

    S. Lanthaler, S. Mishra, G. E. Karniadakis, Error estimates for DeepONets: A deep learning framework in infinite dimensions, Transactions of Mathematics and Its Appli- cations 6 (1) (2022) tnac001. doi:10.1093/imatrm/ tnac001

    work page doi:10.1093/imatrm/ 2022

  24. [24]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhat- tacharya, A. Stuart, A. Anandkumar, Fourier neural opera- tor for parametric partial differential equations, in: Inter- national Conference on Learning Representations (ICLR), 2021, arXiv:2010.08895

    Pith/arXiv arXiv 2021

  25. [25]

    Tripura, S

    T. Tripura, S. Chakraborty, Wavelet neural operator for solving parametric partial differential equations in com- putational mechanics problems, Computer Methods in Applied Mechanics and Engineering 404 (2023) 115783. doi:10.1016/j.cma.2022.115783

    arXiv 2023

  26. [26]

    Cao, Choose a transformer: Fourier or galerkin, in: Advances in Neural Information Processing Sys- tems (NeurIPS), V ol

    S. Cao, Choose a transformer: Fourier or galerkin, in: Advances in Neural Information Processing Sys- tems (NeurIPS), V ol. 34, 2021, pp. 24924–24940, arXiv:2105.14995. 16

    Pith/arXiv arXiv 2021

  27. [27]

    Z. Hao, Z. Wang, H. Su, C. Ying, Y . Dong, S. Liu, Z. Cheng, J. Song, J. Zhu, GNOT: A general neural op- erator transformer for operator learning, in: International Conference on Machine Learning (ICML), V ol. 202 of PMLR, 2023, pp. 12556–12569, arXiv:2302.14376

    Pith/arXiv arXiv 2023

  28. [28]

    M. A. Rahman, Z. E. Ross, K. Azizzadenesheli, U-NO: U- shaped neural operators, Transactions on Machine Learn- ing Research (TMLR), arXiv:2204.11127 (2023)

    Pith/arXiv arXiv 2023

  29. [29]

    Kontolati, S

    K. Kontolati, S. Goswami, G. E. Karniadakis, M. D. Shields, Learning nonlinear operators in latent spaces for real-time predictions of complex dynamics in phys- ical systems, Nature Communications 15 (2024) 5101. doi:10.1038/s41467-024-49411-w

    doi:10.1038/s41467-024-49411-w 2024

  30. [30]

    Mandl, D

    L. Mandl, D. Nayak, T. Ricken, S. Goswami, Physics- informed time-integrated DeepONet: Temporal tangent space operator learning for high-accuracy inference, Com- puter Methods in Applied Mechanics and Engineering 455 (2026) 118917.doi:10.1016/j.cma.2026.118917

    arXiv 2026

  31. [31]

    Buitrago Ruiz, T

    R. Buitrago Ruiz, T. Marwah, A. Gu, A. Risteski, On the benefits of memory for modeling time-dependent PDEs, in: International Conference on Learning Representations (ICLR), 2025, arXiv:2409.02313

    Pith/arXiv arXiv 2025

  32. [32]

    S. Wang, H. Wang, P. Perdikaris, Improved architectures and training algorithms for deep operator networks, Journal of Scientific Computing 92 (2) (2022) 35. doi:10.1007/ s10915-022-01881-0

    2022

  33. [33]

    L. Lu, X. Meng, S. Cai, Z. Mao, S. Goswami, Z. Zhang, G. E. Karniadakis, A comprehensive and fair comparison of two neural operators (with practical extensions) based on FAIR data, Computer Methods in Applied Mechanics and Engineering 393 (2022) 114778. doi:10.1016/j. cma.2022.114778

    arXiv 2022

  34. [34]

    Ye, C.-S

    Z. Ye, C.-S. Zhang, W. Wang, Recurrent neural opera- tors: Stable long-term PDE prediction, arXiv preprint arXiv:2505.20721 (2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Brandstetter, D

    J. Brandstetter, D. E. Worrall, M. Welling, Message passing neural PDE solvers, in: International Conference on Learn- ing Representations (ICLR), 2022, arXiv:2202.03376

    Pith/arXiv arXiv 2022

  36. [36]

    Lippe, B

    P. Lippe, B. S. Veeling, P. Perdikaris, R. E. Turner, J. Brand- stetter, PDE-Refiner: Achieving accurate long rollouts with neural PDE solvers, in: Advances in Neural Information Processing Systems (NeurIPS), V ol. 36, 2023, pp. 67398– 67433, arXiv:2308.05732

    Pith/arXiv arXiv 2023

  37. [37]

    Floryan, On instabilities in neural network-based physics simulators, arXiv preprint arXiv:2406.13101 (2024)

    D. Floryan, On instabilities in neural network-based physics simulators, arXiv preprint arXiv:2406.13101 (2024)

    Pith/arXiv arXiv 2024

  38. [38]

    K. Um, R. Brand, Y . R. Fei, P. Holl, N. Thuerey, Solver-in- the-loop: Learning from differentiable physics to interact with iterative PDE-solvers, in: Advances in Neural Infor- mation Processing Systems (NeurIPS), V ol. 33, 2020, pp. 6111–6122, arXiv:2007.00016

    Pith/arXiv arXiv 2020

  39. [40]

    Clark Di Leoni, L

    P. Clark Di Leoni, L. Lu, C. Meneveau, G. E. Karni- adakis, T. A. Zaki, Neural operator prediction of linear instability waves in high-speed boundary layers, Jour- nal of Computational Physics 474 (2023) 111793. doi: 10.1016/j.jcp.2022.111793

    arXiv 2023

  40. [41]

    S. Wang, P. Perdikaris, Long-time integration of parametric evolution equations with physics-informed DeepONets, Journal of Computational Physics 475 (2023) 111855.doi: 10.1016/j.jcp.2022.111855

    arXiv 2023

  41. [42]

    Nayak, S

    D. Nayak, S. Goswami, TI-DeepONet: Learnable time inte- gration for stable long-term extrapolation, Computer Meth- ods in Applied Mechanics and Engineering 456 (2026) 118960.doi:10.1016/j.cma.2026.118960

    arXiv 2026

  42. [43]

    D. W. Abueidda, M. Nonna, P. Pantidis, M. E. Mobasher, Time resolution independent operator learning, Computer Methods in Applied Mechanics and Engineering 450 (2026) 118586.doi:10.1016/j.cma.2025.118586

    arXiv 2026

  43. [44]

    McGreivy, A

    N. McGreivy, A. Hakim, Weak baselines and report- ing biases lead to overoptimism in machine learning for fluid-related partial differential equations, Nature Machine Intelligence 6 (10) (2024) 1256–1269. doi:10.1038/ s42256-024-00897-5

    2024

  44. [45]

    A. T. Mohan, A. Chattopadhyay, J. Miller, What you see is not what you get: Neural partial differential equations and the illusion of learning, arXiv preprint arXiv:2411.15101 (2024)

    Pith/arXiv arXiv 2024

  45. [46]

    Koehler, N

    F. Koehler, N. Thuerey, Neural emulator superiority: When machine learning for PDEs surpasses its training data, in: Advances in Neural Information Processing Systems (NeurIPS), V ol. 38, 2025, pp. 165660–165702, arXiv:2510.23111

    arXiv 2025

  46. [47]

    X. Zou, Z. Li, Y . Wang, H. Yang, J. Wang, Uncertainty quantification and stability of neural operators for predic- tion of three-dimensional turbulence, Journal of Computa- tional Physics 549 (2026) 114640. doi:10.1016/j.jcp. 2025.114640

    arXiv 2026

  47. [48]

    Westermann, B

    J. Westermann, B. Huber, T. O’Leary-Roseberry, J. Zech, Performance of neural and polynomial operator surrogates, arXiv preprint arXiv:2604.00689 (2026)

    arXiv 2026

  48. [49]

    I. A. Baratta, J. P. Dean, J. S. Dokken, M. Habera, J. S. Hale, C. N. Richardson, M. E. Rognes, M. W. Scroggs, N. Sime, G. N. Wells, DOLFINx: The next generation FEniCS problem solving environment, zenodo (2023). doi:10.5281/zenodo.10447666. 17

    doi:10.5281/zenodo.10447666 2023

  49. [50]

    F. W. Preston, The theory and design of plate glass polish- ing machines, Journal of the Society of Glass Technology 11 (1927) 214–256

    1927

  50. [51]

    Borucki, Y

    L. Borucki, Y . Sampurno, A. Philipossian, The shear force law: A guide to modeling CMP removal rates, ECS Jour- nal of Solid State Science and Technology 12 (4) (2023) 044003.doi:10.1149/2162-8777/accaa6

    work page doi:10.1149/2162-8777/accaa6 2023

  51. [52]

    C. H. K. Williamson, V ortex dynamics in the cylinder wake, Annual Review of Fluid Mechanics 28 (1996) 477–539. doi:10.1146/annurev.fl.28.010196.002401

    arXiv 1996

  52. [53]

    Schäfer, S

    M. Schäfer, S. Turek, F. Durst, E. Krause, R. Rannacher, Benchmark computations of laminar flow around a cylin- der, in: E. H. Hirschel (Ed.), Flow Simulation with High- Performance Computers II, V ol. 48 of Notes on Numerical Fluid Mechanics, Vieweg+Teubner Verlag, 1996, pp. 547– 566.doi:10.1007/978-3-322-89849-4_39

    doi:10.1007/978-3-322-89849-4_39 1996

  53. [54]

    J. L. Proctor, S. L. Brunton, J. N. Kutz, Dynamic mode decomposition with control, SIAM Journal on Applied Dy- namical Systems 15 (1) (2016) 142–161. doi:10.1137/ 15M1013857

    2016

  54. [55]

    S. L. Brunton, B. W. Brunton, J. L. Proctor, J. N. Kutz, Koopman invariant subspaces and finite linear representa- tions of nonlinear dynamical systems for control, PLOS ONE 11 (2) (2016) e0150171. doi:10.1371/journal. pone.0150171

    doi:10.1371/journal 2016

  55. [56]

    Lusch, J

    B. Lusch, J. N. Kutz, S. L. Brunton, Deep learning for universal linear embeddings of nonlinear dynamics, Na- ture Communications 9 (2018) 4950. doi:10.1038/ s41467-018-07210-0

    2018

  56. [58]

    D. E. Rumelhart, G. E. Hinton, R. J. Williams, Learn- ing representations by back-propagating errors, Nature 323 (6088) (1986) 533–536.doi:10.1038/323533a0

    doi:10.1038/323533a0 1986

  57. [59]

    J. He, S. Kushwaha, J. Park, S. Koric, D. Abueidda, I. Ja- siuk, Sequential deep operator networks (S-DeepONet) for predicting full-field solutions under time-dependent loads, Engineering Applications of Artificial Intelligence 127 (2024) 107258. doi:10.1016/j.engappai.2023. 107258

    doi:10.1016/j.engappai.2023 2024

  58. [60]

    K. Cho, B. van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, Y . Bengio, Learning phrase representations using RNN encoder–decoder for statistical machine translation, in: Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014, pp. 1724– 1734.doi:10.3115/v1/D14-1179

    doi:10.3115/v1/d14-1179 2014

  59. [61]

    A. Gu, K. Goel, C. Ré, Efficiently modeling long se- quences with structured state spaces, in: International Conference on Learning Representations (ICLR), 2022, arXiv:2111.00396

    Pith/arXiv arXiv 2022

  60. [62]

    A. Gu, K. Goel, A. Gupta, C. Ré, On the parameter- ization and initialization of diagonal state space mod- els, in: Advances in Neural Information Processing Systems (NeurIPS), V ol. 35, 2022, pp. 35971–35983, arXiv:2206.11893

    Pith/arXiv arXiv 2022

  61. [63]

    A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces, in: First Conference on Lan- guage Modeling (COLM), 2024, arXiv:2312.00752

    Pith/arXiv arXiv 2024

  62. [64]

    Gu, Modeling sequences with structured state spaces, Ph.D

    A. Gu, Modeling sequences with structured state spaces, Ph.D. thesis, Stanford University, https://purl.stanford.edu/mb976vf9362 (2023)

    2023

  63. [65]

    M. H. Baattrup, J. Bach, L. Jeppe, F. Labe, A. Grohsjean, C. Schwanenberger, P. Stelldinger, Pointwise metrics mis- lead: An evaluation protocol for multimodal inverse prob- lems, arXiv preprint arXiv:2605.22891 (2026)

    Pith/arXiv arXiv 2026

  64. [66]

    Gelman, E

    A. Gelman, E. Loken, The statistical crisis in science, American Scientist 102 (6) (2014) 460–465. doi:10. 1511/2014.111.460

    2014

  65. [67]

    N. R. Franco, A. Manzoni, P. Zunino, J. S. Hesthaven, Deep orthogonal decomposition: A continuously adap- tive neural network approach to model order reduction of parametrized partial differential equations, Advances in Computational Mathematics 52 (3) (2026) 31. doi: 10.1007/s10444-026-10295-7

    work page doi:10.1007/s10444-026-10295-7 2026

  66. [68]

    Y . Qiu, W. Dahmen, P. Chen, Variationally correct operator learning: Reduced basis neural operator with a posteriori error estimation, Computer Methods in Applied Mechanics and Engineering 461 (2026) 119228. doi:10.1016/j. cma.2026.119228

    arXiv 2026

  67. [69]

    Z. Li, N. B. Kovachki, C. Choy, B. Li, J. Kossaifi, S. P. Otta, M. A. Nabian, M. Stadler, C. Hundt, K. Azizzadenesheli, A. Anandkumar, Geometry-informed neural operator for large-scale 3D PDEs, in: Advances in Neural Information Processing Systems (NeurIPS), V ol. 36, 2023, pp. 35836– 35854, arXiv:2309.00583

    Pith/arXiv arXiv 2023

  68. [70]

    Peyvan, V

    A. Peyvan, V . Kumar, G. E. Karniadakis, Fusion- DeepONet: A data-efficient neural operator for geometry- dependent hypersonic and supersonic flows, Journal of Computational Physics 544 (2026) 114432. doi:10. 1016/j.jcp.2025.114432

    arXiv 2026

  69. [71]

    Musekamp, M

    D. Musekamp, M. Kalimuthu, D. Holzmüller, M. Takamoto, M. Niepert, Active learning for neu- ral PDE solvers, in: International Conference on Learning Representations (ICLR), 2025, arXiv:2408.01536. 18

    Pith/arXiv arXiv 2025

  70. [72]

    Huang, D

    Y . Huang, D. S. Greenberg, Geometric and physical con- straints synergistically enhance neural PDE surrogates, in: Proceedings of the 42nd International Conference on Ma- chine Learning (ICML), 2025, arXiv:2506.05513. 19

    Pith/arXiv arXiv 2025