arxiv: 2605.13912 · v1 · submitted 2026-05-13 · 🧮 math.NA · cs.NA· physics.flu-dyn

Recognition: no theorem link

ViT-K: A Few-Shot Learning Model for Coupled Fluid-Porous Media Flows with Interface Conditions

Mengjia Chen , Changxin Qiu , Zhiping Mao , Menghui Xu

Authors on Pith no claims yet

Pith reviewed 2026-05-15 02:55 UTC · model grok-4.3

classification 🧮 math.NA cs.NAphysics.flu-dyn

keywords few-shot learningKoopman operatorVision Transformercoupled fluid-porous flowsinterface conditionsNavier-Stokes-Darcysurrogate modelinglong-term prediction

0 comments

The pith

ViT-K learns stable long-term predictions for coupled fluid-porous flows from few examples by linearizing dynamics with a Koopman operator.

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

The paper introduces ViT-K as a few-shot learning model for the interaction between free flow and porous media under coupled Stokes, Navier-Stokes, and Darcy equations with interface conditions. It pairs vision transformers to extract heterogeneous features at the interfaces with the Koopman operator to lift the nonlinear evolution onto a linear observable space. This construction produces stability by design, so that forecast errors accumulate only linearly with time rather than exponentially. The result matters for applications that require fast, reliable forecasts of filtration and transport processes where conventional grid-based solvers become prohibitive over long horizons or with limited training data.

Core claim

By lifting nonlinear dynamics into a globally linear observable space, the ViT-K model provides stability by design, ensuring that prediction errors grow linearly rather than exponentially over time. This theoretical property enables reliable long-term extrapolation even in small-sample regimes. Numerical experiments on benchmark coupled systems show that the approach reconstructs interface physics with high fidelity and remains robust to measurement noise through implicit spectral filtering.

What carries the argument

The ViT-K framework that fuses Vision Transformers for spatial capture of heterogeneous interfacial features with the Koopman operator for global linearization of the temporal dynamics on a low-dimensional manifold.

If this is right

ViT-K reconstructs complex interface physics from sparse data with high fidelity on benchmark coupled systems.
The model acts as an implicit spectral filter, conferring robustness to measurement noise.
Inference speed exceeds that of traditional solvers while physical consistency is preserved.
Long-term forecasts remain reliable because errors accumulate only linearly in time.

Where Pith is reading between the lines

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

The same transformer-plus-Koopman structure could be tested on other sharp-interface multi-physics problems such as free-surface flows or biological transport.
Because the linearization supplies built-in stability, the model may integrate directly with model-predictive control schemes for real-time flow management.
Scaling the architecture to three-dimensional domains or time-varying interfaces would provide a direct test of whether the linear error growth persists beyond the reported benchmarks.

Load-bearing premise

The Koopman operator can be learned from sparse data to linearize the full coupled Stokes-Navier-Stokes-Darcy system including interface conditions without discarding essential nonlinear behavior or demanding heavy hyperparameter tuning.

What would settle it

A controlled experiment on the standard benchmark coupled flow problems in which long-term prediction errors of ViT-K are shown to grow exponentially rather than linearly would refute the stability-by-design claim.

Figures

Figures reproduced from arXiv: 2605.13912 by Changxin Qiu, Menghui Xu, Mengjia Chen, Zhiping Mao.

**Figure 1.** Figure 1: The ViT-K Framework Architecture. 2. Model Problems. Consider a bounded domain Ω ⊂ R d (d ∈ {2, 3}) partitioned into a porous medium region ΩD and a free-flow fluid region ΩS, separated by a common interface Γ = ΩD ∩ ΩS. Let ⃗nS and ⃗nD denote the unit outward normal vectors on Γ for ΩS and ΩD, respectively. Γ Γ [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Schematic representation of the computational domain Ω, consisting of the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Schematic of the standard Vision Transformer architecture. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Architecture of the proposed ViT-K framework. The ViT encoder extracts [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Root Mean Square Error (RMSE) vs. Extrapolation Time [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

**Figure 6.** Figure 6: Convergence history of the component-wise training losses, showing stable [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

**Figure 7.** Figure 7: Temporal evolution of velocity u1 at (0.5, −0.08) using ViT-K, FNO, and ConvLSTM [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

**Figure 8.** Figure 8: A sketch of flow in a Karst aquifer with curved interfaces for Example 3. uS =    (ω1, 0), on HA (Inlet 1), (0, ω1), on CD (Inlet 2), (ω2, 0), on F G (Outlet). Dataset and Configuration: Reference solutions are computed via a highfidelity finite element method (FEM) solver using P2-P1 elements for uS, p and P2 for ϕD. The training set utilizes sparse temporal snapshots at t ∈ {0.2, 0.4, 0.6, 0.8}, r… view at source ↗

**Figure 9.** Figure 9: Predictive performance on the Karst aquifer model at extrapolation time [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗

**Figure 10.** Figure 10: Comparison at t = 1.5 s (Extrapolation Phase). Long-term stability is maintained due to the non-autonomous forcing formulation. Reference (FEM) Prediction (ViT-K) 0.0 0.2 0.4 0.6 0.8 1.0 Pressure field 0.0 0.2 0.4 0.6 0.8 1.0 Velocity magnitude (a) t = 0.5 s Reference (FEM) Prediction (ViT-K) 0.0 0.1 0.2 0.3 0.4 0.5 Pressure field 0.0 0.2 0.4 0.6 0.8 Velocity magnitude (b) t = 1.0 s Reference (FEM) Predi… view at source ↗

**Figure 11.** Figure 11: Evolution of flow field streamlines and pressure over multiple cardiac cycles. [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗

**Figure 12.** Figure 12: tracks the Root Mean Squared Error (RMSE) evolution over the extreme extrapolation horizon up to t = 100.0 s. Crucially, the phase-averaged error follows a strictly linear trajectory with a slope of 10−4 . This empirical result perfectly corroborates the theoretical bound derived in Theorem 4.2, confirming that the stable Koopman formulation successfully suppresses exponential divergence even when extra… view at source ↗

**Figure 13.** Figure 13: Long-term flow predictions for the bifurcating vessel. ViT-K preserves pul [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗

**Figure 14.** Figure 14: Comparison of the FEM reference with Gaussian noise (left column), the ViT [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗

**Figure 15.** Figure 15: Prediction performance of physical quantity [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗

read the original abstract

The numerical simulation of interaction between free flow and porous media, governed by coupled Stokes/Navier--Stokes--Darcy flows, is critical for understanding fluid filtration and physiological transport, yet it is hindered by the high computational cost of resolving interface heterogeneities and the instability of long-term predictions. While deep learning offers surrogate modeling potential, existing frameworks often suffer from exponential error accumulation and poor convergence in multi-physics regimes. To address these limitations, we propose ViT-K, a novel few-shot learning model designed to learn the spatiotemporal evolution of coupled flows from sparse datasets. The ViT-K framework effectively reconstructs the global flow physics on a low-dimensional manifold by combining Vision Transformers (ViT) to capture heterogeneous interfacial features with the Koopman operator to linearize temporal dynamics. By lifting nonlinear dynamics into a globally linear observable space, the ViT-K model provides stability by design, ensuring that prediction errors grow linearly rather than exponentially over time. This theoretical property enables reliable long-term extrapolation even in small-sample regimes. Numerical experiments on benchmark coupled systems demonstrate that ViT-K not only captures complex interface physics with high fidelity but also exhibits exceptional robustness against measurement noise by acting as an implicit spectral filter. The proposed method significantly outperforms traditional solvers in inference speed while maintaining physical consistency, offering a robust paradigm for real-time multiphysics forecasting.

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.

Desk Editor's Note private letter to a colleague

ViT-K pairs ViT interface handling with Koopman linearization for few-shot coupled flow prediction, but the linear-error-growth guarantee is asserted rather than demonstrated from the sparse-data operator.

read the letter

The main point is a hybrid architecture that uses Vision Transformers to extract heterogeneous features at the free-flow/porous interface and Koopman lifting to turn the time-stepping into a linear map, aimed at stable extrapolation from small datasets in Stokes/Navier-Stokes-Darcy systems. The paper correctly flags the practical cost of resolving interface conditions and the risk of exponential error buildup in long rollouts, and it tries to solve both with one model. That pairing is not a standard extension of either ViT or Koopman work in this exact multiphysics setting, so the combination itself is the concrete novelty. The abstract also notes noise robustness and faster inference than traditional solvers, which would be useful if the numbers hold. The soft spot is the stability claim. It rests on the learned Koopman operator producing a globally linear observable space that exactly reproduces the coupled dynamics, including the interface jumps. Because the operator comes from regression on sparse samples, any approximation error in the ViT-extracted features or in the chosen observables leaves a residual. In the advective free-flow region that residual can still produce local exponential divergence, and the abstract gives no a-priori bound or spectral check on the learned operator to rule this out. Without the methods section, data splits, error-growth plots, or ablation on observable dimension, the central linear-growth assertion stays unverified. The experiments are described only at the level of “high fidelity and noise robustness,” so it is impossible to judge how well the claim survives on the actual benchmarks. This paper is for people already working on scientific machine learning for multiphysics CFD who want to see a new hybrid tried on interface problems. A reader who needs a working surrogate right now will not get enough implementation detail, but someone looking for architecture ideas in few-shot regimes could extract a useful direction. It deserves peer review because the problem is real, the proposed fix is coherent on paper, and the gaps are fixable with added analysis rather than fundamental contradictions.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces ViT-K, a few-shot learning framework combining Vision Transformers to capture heterogeneous interfacial features with the Koopman operator to linearize the spatiotemporal evolution of coupled Stokes/Navier-Stokes-Darcy flows subject to interface conditions. It claims that lifting the nonlinear dynamics into a globally linear observable space yields stability by design, so that prediction errors grow linearly (rather than exponentially) over time, enabling reliable long-term extrapolation from sparse data. Numerical experiments on benchmark coupled systems are reported to demonstrate high-fidelity reconstruction of interface physics, exceptional noise robustness, and faster inference than traditional solvers while preserving physical consistency.

Significance. If the central stability claim holds, the work would constitute a meaningful contribution to surrogate modeling of multiphysics flows by supplying a theoretically grounded, data-efficient method for long-term forecasting. The explicit handling of interface jump conditions via ViT and the Koopman linearization together address both computational cost and instability, with clear relevance to filtration and physiological transport applications. The few-shot regime and implicit spectral filtering are practically attractive strengths.

major comments (2)

[Abstract] Abstract: the claim that 'lifting nonlinear dynamics into a globally linear observable space... ensures that prediction errors grow linearly rather than exponentially' is not accompanied by an a-priori residual bound or spectral-radius analysis of the learned Koopman operator; because the operator is obtained by regression on sparse samples, any approximation error in the ViT-extracted interface features can leave a nonzero residual that permits local exponential divergence in the advective free-flow region, directly undermining the linear-error-growth guarantee.
[Numerical experiments] Numerical experiments section: the reported high-fidelity and noise-robust results are presented without data-split details, error bars, ablation studies on observable dimension or patch size, or quantitative comparison of long-term error growth rates against a non-Koopman baseline; this leaves the central linear-error-growth claim unverified at the level required for the stability-by-design assertion.

minor comments (2)

The free parameters (Koopman observable dimension, ViT patch size, embedding dimension) should be tabulated with the specific values used in each benchmark experiment.
Notation for the interface jump conditions and the precise form of the Koopman observables should be introduced explicitly before the stability argument is invoked.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the insightful and constructive comments. We address each major point below and will incorporate revisions to strengthen the theoretical and experimental support for the stability claims.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that 'lifting nonlinear dynamics into a globally linear observable space... ensures that prediction errors grow linearly rather than exponentially' is not accompanied by an a-priori residual bound or spectral-radius analysis of the learned Koopman operator; because the operator is obtained by regression on sparse samples, any approximation error in the ViT-extracted interface features can leave a nonzero residual that permits local exponential divergence in the advective free-flow region, directly undermining the linear-error-growth guarantee.

Authors: We agree that the linear-error-growth property is rigorously guaranteed only for the exact Koopman operator with spectral radius at most one. For the learned operator obtained via regression on finite samples, approximation errors in the ViT features can indeed introduce a nonzero residual. In the revised manuscript we will add a dedicated subsection deriving a residual bound that accounts for both the finite-data regression error and the ViT approximation error, together with the computed spectral radius of the learned operator on the training trajectories. This analysis will explicitly state the conditions under which the linear growth regime is preserved. revision: yes
Referee: [Numerical experiments] Numerical experiments section: the reported high-fidelity and noise-robust results are presented without data-split details, error bars, ablation studies on observable dimension or patch size, or quantitative comparison of long-term error growth rates against a non-Koopman baseline; this leaves the central linear-error-growth claim unverified at the level required for the stability-by-design assertion.

Authors: We acknowledge that the current experimental section lacks the quantitative details needed to fully substantiate the stability claim. The revised version will include: explicit train/validation/test split ratios and random-seed information; error bars computed over five independent runs; ablation tables varying observable dimension and ViT patch size with corresponding long-term prediction metrics; and a direct side-by-side comparison of error-growth curves against a non-Koopman baseline (LSTM surrogate) over 500 time steps. These additions will provide the requested verification of linear versus exponential error accumulation. revision: yes

Circularity Check

0 steps flagged

No significant circularity: stability claim follows from standard Koopman property independent of fitted values

full rationale

The paper's derivation chain combines ViT feature extraction with the Koopman operator to produce a linear model in the lifted space. The asserted linear error growth is a direct mathematical consequence of applying a linear operator to observables, which holds by construction of the Koopman framework itself and does not reduce to any fitted parameter or self-citation within the paper. No equation equates a prediction to its own training target by definition, and the few-shot regime is treated as an empirical regime rather than a definitional input. The interface conditions are handled via the ViT component, but this does not create a self-referential loop in the stability argument. The derivation remains self-contained against external dynamical-systems benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim depends on the domain assumption that Koopman observables exist and can be learned from few-shot data for this interface problem, plus standard neural-network training assumptions.

free parameters (2)

Koopman observable dimension and basis functions
Chosen or learned to achieve linearization; directly affects the stability guarantee.
ViT patch size and embedding dimension
Hyperparameters fitted during few-shot training on interface data.

axioms (1)

domain assumption Nonlinear coupled flow dynamics admit a finite-dimensional Koopman linearization from sparse observations
Invoked to justify linear error growth; appears in the description of lifting to observable space.

invented entities (1)

ViT-K framework no independent evidence
purpose: Reconstructs global flow physics on low-dimensional manifold from sparse data
New composite model introduced by the paper; no independent evidence provided beyond the reported experiments.

pith-pipeline@v0.9.0 · 5557 in / 1344 out tokens · 52261 ms · 2026-05-15T02:55:15.665863+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages

[1]

Advances in neural information processing systems , volume=

Convolutional LSTM network: A machine learning approach for precipitation nowcasting , author=. Advances in neural information processing systems , volume=

work page
[2]

Journal of Hydrology , volume=

A fully coupled depth-integrated model for surface water and groundwater flows , author=. Journal of Hydrology , volume=. 2016 , publisher=

work page 2016
[3]

2004 , type=

Domain decomposition methods for the coupling of surface and groundwater flows , author=. 2004 , type=

work page 2004
[4]

Calcolo , volume=

Computational issues related to iterative coupling of subsurface and channel flows , author=. Calcolo , volume=. 2007 , publisher=

work page 2007
[5]

Transport in porous media , volume=

Numerical analysis of coupled Stokes/Darcy flows in industrial filtrations , author=. Transport in porous media , volume=. 2006 , publisher=

work page 2006
[6]

Lab on a Chip , volume=

Fabrication and verification of a glass--silicon--glass micro-/nanofluidic model for investigating multi-phase flow in shale-like unconventional dual-porosity tight porous media , author=. Lab on a Chip , volume=. 2019 , publisher=

work page 2019
[7]

Scientific Reports , volume=

A comprehensive mathematical model for cardiac perfusion , author=. Scientific Reports , volume=. 2023 , publisher=

work page 2023
[8]

International Journal of Heat and Mass Transfer , volume=

The role of porous media in modeling flow and heat transfer in biological tissues , author=. International Journal of Heat and Mass Transfer , volume=. 2003 , publisher=

work page 2003
[9]

SIAM Journal on Numerical Analysis , volume=

A robust finite element method for Darcy--Stokes flow , author=. SIAM Journal on Numerical Analysis , volume=. 2002 , publisher=

work page 2002
[10]

SIAM Journal on Numerical Analysis , volume=

Coupling fluid flow with porous media flow , author=. SIAM Journal on Numerical Analysis , volume=. 2002 , publisher=

work page 2002
[11]

Computational geosciences , volume=

A computational method for approximating a Darcy--Stokes system governing a vuggy porous medium , author=. Computational geosciences , volume=. 2007 , publisher=

work page 2007
[12]

SIAM Journal on Numerical Analysis , volume=

Finite element approximations for Stokes--Darcy flow with Beavers--Joseph interface conditions , author=. SIAM Journal on Numerical Analysis , volume=. 2010 , publisher=

work page 2010
[13]

Computer Methods in Applied Mechanics and Engineering , volume=

New fully-mixed finite element methods for the Stokes--Darcy coupling , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2015 , publisher=

work page 2015
[14]

IMA Journal of Numerical Analysis , volume=

A conforming mixed finite-element method for the coupling of fluid flow with porous media flow , author=. IMA Journal of Numerical Analysis , volume=. 2009 , publisher=

work page 2009
[15]

Applied Numerical Mathematics , volume=

A stabilized finite volume element method for a coupled Stokes--Darcy problem , author=. Applied Numerical Mathematics , volume=. 2018 , publisher=

work page 2018
[16]

Numerische Mathematik , volume=

Discontinuous Galerkin and mimetic finite difference methods for coupled Stokes--Darcy flows on polygonal and polyhedral grids , author=. Numerische Mathematik , volume=. 2014 , publisher=

work page 2014
[17]

SIAM Journal on Numerical Analysis , volume=

Locally conservative coupling of Stokes and Darcy flows , author=. SIAM Journal on Numerical Analysis , volume=. 2005 , publisher=

work page 2005
[18]

ESAIM: Mathematical Modelling and Numerical Analysis , volume=

Mortar finite element discretization of a model coupling Darcy and Stokes equations , author=. ESAIM: Mathematical Modelling and Numerical Analysis , volume=. 2008 , publisher=

work page 2008
[19]

Communications in Applied Mathematics and Computational Science , volume=

FETI and BDD preconditioners for Stokes--Mortar--Darcy systems , author=. Communications in Applied Mathematics and Computational Science , volume=. 2009 , publisher=

work page 2009
[20]

SIAM Journal on Numerical Analysis , volume=

Robin--Robin domain decomposition methods for the Stokes--Darcy coupling , author=. SIAM Journal on Numerical Analysis , volume=. 2007 , publisher=

work page 2007
[21]

Journal of Computational Physics , volume=

A domain decomposition method for the time-dependent Navier-Stokes-Darcy model with Beavers-Joseph interface condition and defective boundary condition , author=. Journal of Computational Physics , volume=. 2020 , publisher=

work page 2020
[22]

SIAM journal on numerical analysis , volume=

A two-grid method of a mixed Stokes--Darcy model for coupling fluid flow with porous media flow , author=. SIAM journal on numerical analysis , volume=. 2007 , publisher=

work page 2007
[23]

Acta Mechanica Sinica , volume=

Physics-informed neural networks (PINNs) for fluid mechanics: A review , author=. Acta Mechanica Sinica , volume=. 2021 , publisher=

work page 2021
[24]

2020 , author =

Physics-informed neural networks for high-speed flows , journal =. 2020 , author =

work page 2020
[25]

Entropy , volume=

Physics-informed neural networks for solving coupled Stokes--Darcy equation , author=. Entropy , volume=. 2022 , publisher=

work page 2022
[26]

Applied Mathematics and Computation , volume=

Efficient coupled deep neural networks for the time-dependent coupled Stokes-Darcy problems , author=. Applied Mathematics and Computation , volume=. 2023 , publisher=

work page 2023
[27]

Computers & Mathematics with Applications , volume=

MC-CDNNs: The Monte Carlo-coupled deep neural networks approach for stochastic dual-porosity-Stokes flow coupled model , author=. Computers & Mathematics with Applications , volume=. 2025 , publisher=

work page 2025
[28]

Entropy , volume=

High-Accuracy Parallel Neural Networks with Hard Constraints for a Mixed Stokes/Darcy Model , author=. Entropy , volume=. 2025 , publisher=

work page 2025
[29]

International Conference on Learning Representations (ICLR) , year=

Fourier Neural Operator for Parametric Partial Differential Equations , author=. International Conference on Learning Representations (ICLR) , year=

work page
[30]

Nature machine intelligence , volume=

Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators , author=. Nature machine intelligence , volume=. 2021 , publisher=

work page 2021
[31]

Rahman, Md Ashiqur and Ross, Zachary E and Azizzadenesheli, Kamyar , journal=

work page
[32]

ACM/IMS Journal of Data Science , volume=

Physics-informed neural operator for learning partial differential equations , author=. ACM/IMS Journal of Data Science , volume=. 2024 , publisher=

work page 2024
[33]

Advances in neural information processing systems , volume=

Characterizing possible failure modes in physics-informed neural networks , author=. Advances in neural information processing systems , volume=

work page
[34]

Advances in Neural Information Processing Systems , volume=

Pdebench: An extensive benchmark for scientific machine learning , author=. Advances in Neural Information Processing Systems , volume=

work page
[35]

Advances in Neural Information Processing Systems , volume=

Pde-refiner: Achieving accurate long rollouts with neural pde solvers , author=. Advances in Neural Information Processing Systems , volume=

work page
[36]

Computer Methods in Applied Mechanics and Engineering , volume=

Time-marching neural operator--FE coupling: AI-accelerated physics modeling , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2025 , publisher=

work page 2025
[37]

AIP Advances , volume=

A new fluid flow approximation method using a vision transformer and a U-shaped convolutional neural network , author=. AIP Advances , volume=. 2023 , publisher=

work page 2023
[38]

Ocean Engineering , volume=

Twins-PIVNet: Spatial attention-based deep learning framework for particle image velocimetry using Vision Transformer , author=. Ocean Engineering , volume=. 2025 , publisher=

work page 2025
[39]

Brunton, Steven L and Budi. Modern. SIAM Review , volume=. 2022 , publisher=

work page 2022
[40]

Journal of fluid mechanics , volume=

Boundary conditions at a naturally permeable wall , author=. Journal of fluid mechanics , volume=. 1967 , publisher=

work page 1967
[41]

Journal of Nonlinear Science , volume =

A Data--Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition , author =. Journal of Nonlinear Science , volume =. 2015 , doi =

work page 2015
[42]

International Conference on Learning Representations (ICLR) , year=

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations (ICLR) , year=

work page
[43]

Advances in neural information processing systems , volume=

Attention is all you need , author=. Advances in neural information processing systems , volume=

work page
[44]

International conference on machine learning , pages=

Training data-efficient image transformers & distillation through attention , author=. International conference on machine learning , pages=. 2021 , organization=

work page 2021
[45]

Proceedings of the National Academy of Sciences , volume=

Hamiltonian systems and transformation in Hilbert space , author=. Proceedings of the National Academy of Sciences , volume=

work page
[46]

Nonlinear Dynamics , volume=

Spectral properties of dynamical systems, model reduction and decompositions , author=. Nonlinear Dynamics , volume=. 2005 , publisher=

work page 2005
[47]

Automatica , volume=

Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control , author=. Automatica , volume=. 2018 , publisher=

work page 2018
[48]

International Conference on Learning Representations (ICLR) , year=

Are Transformers universal approximators of sequence-to-sequence functions? , author=. International Conference on Learning Representations (ICLR) , year=

work page
[49]

IEEE Transactions on Information theory , volume=

Universal approximation bounds for superpositions of a sigmoidal function , author=. IEEE Transactions on Information theory , volume=. 2002 , publisher=

work page 2002
[50]

Journal of Machine Learning Research , volume=

Optimal approximation rates for deep ReLU neural networks on Sobolev and Besov spaces , author=. Journal of Machine Learning Research , volume=

work page
[51]

Journal of machine learning research , volume=

Rademacher and gaussian complexities: Risk bounds and structural results , author=. Journal of machine learning research , volume=

work page
[52]

Applied Mechanics Reviews , volume=

Reynolds-averaged Navier-Stokes equations for turbulence modeling , author=. Applied Mechanics Reviews , volume=. 2009 , publisher=

work page 2009
[53]

arXiv preprint arXiv:2405.11223 , year=

A class of new linear, efficient and high-order implicit-explicit methods for the coupled free flow-porous media system based on nonlinear Lions interface condition , author=. arXiv preprint arXiv:2405.11223 , year=

work page arXiv