betaFeedback

WorksComputer Science

NAGO: Noise-Aware Generative Operator via Flow Matching

Abstract

We introduce NAGO, a noise-aware generative operator for efficient sampling from invariant measures of stochastic partial differential equations (SPDEs) in function space. Our central idea is to treat the source measure as a design variable rather than a fixed Gaussian default. We establish admissibility conditions for well-defined infinite-dimensional flows and show that the finite-band source–target covariance ratio controls the spectral stiffness of the marginal velocity field. This analysis yields an admissibility-constrained spectral-matching rule: choose a source that is sufficiently rough for well-posedness, then minimize covariance mismatch over the resolved Fourier modes. On three representative stochastic-field benchmarks, this source design improves few-step spectral accuracy and several complementary distributional and physical statistics. In a controlled Matern-field ablation, it reduces the accurate RK4 budget from about 30 steps for Gaussian functional flow matching to 3–5 steps, a 66–10×10\times reduction in neural ODE evaluations. These gains are strongest for stationary multiscale fields and depend on the metric, resolution, and guidance protocol.

Introduction

Stochastic partial differential equations (SPDEs) constitute a fundamental mathematical framework for modeling the evolution of infinite-dimensional random systems under uncertainty and provide a rigorous description of diverse spatiotemporal phenomena, including statistical physics [27] and composite materials [23]. However, in many practical applications, such as geophysical modeling [37], turbulent flow [30, 20, 5], and thin film growth [19], the exact evolution of a single realization is often less important than the long-term statistical behavior of the system [13].

A central problem in this context is the characterization of invariant measures [11, 12], which represent the most compressed yet complete description of long-term stochastic behavior, as discussed in Appendix § 6.2. For many SPDEs of physical and statistical interest, these invariant measures take the form of Gaussian or non-Gaussian probability measures on suitable function spaces, with representative examples including the Ornstein–Uhlenbeck equations [52] and nonlinear models such as the stochastic Allen–Cahn equation [58] and the stochastic Navier–Stokes equation [3], which may admit unique ergodic invariant measures under appropriate conditions. Despite their importance, sampling invariant measures of SPDEs remains challenging due to the infinite-dimensional nature of the dynamics, the absence of closed-form solutions in nonlinear regimes, and the intricate interaction between noise, dissipation, and spatial operators. Classical numerical approaches, including long-time integration of discretized SPDEs [42, 12] and Markov chain Monte Carlo (MCMC) sampling in function space [8], may approximate invariant measures but are often computationally prohibitive, particularly in high-dimensional or stiff systems exhibiting pronounced multiscale structure.

These challenges have motivated growing interest in data-driven and generative modeling approaches that aim to learn invariant measures directly, bypassing explicit long-time simulation. Recent advances in generative modeling [22, 56 Lipman et al.] and operator-learning frameworks [33, 35] have shown considerable promise: generative models provide expressive distributional approximations, while neural operators enable mappings between functional distributions. Recent progress has been made toward generative modeling directly in Hilbert or Banach spaces. Rahman et al. [ [49], Seidman et al. [ [53] proposed GANO and VANO, while Pidstrigach et al. [ [48], Lim et al. [ [36], Kerrigan et al. [ [28, 29], Franzese et al. [ [15] develop the theoretical foundations for diffusion models and flow matching in infinite-dimensional settings. More details about related work are reviewed in Appendix § 6.1.

Nevertheless, an important gap remains between the theory of functional generative modeling and its practical use for multiscale SPDE data. Most implementations still adopt standard Gaussian noise as the source distribution by default. For invariant measures with nontrivial spatial regularity, this choice can be poorly conditioned: the discretized Gaussian noise injects excessive high-frequency energy, forcing the learned velocity field to suppress unresolved modes and thereby making training and inference resolution-dependent. Existing remedies, such as interpolation schedules [9], Schrödinger bridge formulations [54], and adaptive numerical solvers [10], can improve the transport dynamics or numerical integration, but they do not directly modify the source measure that induces the stiffness. We address this issue by treating the source measure as a design variable in functional flow matching. This perspective leads to an admissibility-constrained spectral-matching principle: the source must first be rough enough for function-space well-posedness, and should then be chosen to reduce covariance mismatch over the frequencies resolved by the model.

Based on this principle, we propose NAGO, a noise-designed generative operator for sampling from multiscale SPDE invariant measures. By selecting an admissible Matérn source with a finite-band spectrum aligned to the target, NAGO reduces high-frequency mismatch and yields better-conditioned marginal velocity fields. We support this approach with an infinite-dimensional analysis, a quantitative link between source–target covariance ratios and spectral stiffness, and an operational source-selection rule. Our main contributions are summarized as follows:

  • We identify a function-space mechanism behind source design: source regularity determines admissibility, while the resolved source–target covariance ratio governs spectral stiffness.

  • We develop an admissibility-constrained spectral-matching rule for selecting Matérn sources from target samples, including both smoothness and correlation-length selection.

  • Across three stochastic-field benchmarks, we demonstrate improved few-step generation and complementary distributional statistics. A controlled ablation reduces the accurate RK4 budget by 6–10× relative to a Gaussian-source model; a separate post-training benchmark shows an 89.1× raw-throughput ratio to the SNS solver.

Code availability. The code release is available at https://github.com/cgymmy/NAGO.

Preliminary

In this section, the fundamental concepts of random fields and SPDEs are first introduced, followed by a discussion of invariant measures of SPDEs, and concluded with a brief overview of the basic theory underlying functional flow-matching generative models.

Random Field and Stochastic PDEs

To formalize invariant measures and noise distributions in function space, we briefly recall the notion of random fields. Usually, a random field is regarded as a random variable taking values in a function space; we formalize this as follows:

Definition 2.1 (Random Field). For a set D⊂RdD \subset\mathbb{R}^{d}, a (real-valued) random field {u(x):x∈D}\{u(x): x \in D\} is a collection of real-valued random variables defined on a probability space (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}), where Ω\Omega denotes the sample functional space, F\mathcal{F} is a σ\sigma-algebra of subsets of Ω\Omega, and P\mathbb{P} is a probability measure on (Ω,F)(\Omega,\mathcal{F}).

The class of random fields is broad and contains many important special cases with distinct structural properties as discussed in Appendix C.2.4. Among them, the Matérn field [37] plays a particularly prominent role, as it defines a Gaussian random field with tunable smoothness and correlation length. These properties make it both mathematically tractable and practically useful, especially for studying the effect of regularity in function-space probabilistic modeling. We therefore introduce the Matérn field as a representative example.

Definition 2.2 (Matérn Field). A Gaussian random field u(x)u(x) with Matérn covariance is called a Matérn field, which is a solution to the linear fractional stochastic partial differential equation:

Lα/2u(x)=(κ2−Δ)α/2u(x)=W(x),(1)\mathcal{L}^{\alpha/2}u(x) = \left(\kappa^{2} - \Delta\right)^{\alpha/2}u(x) = W(x), \tag*{(1)}

where x∈D⊂Rdx \in D \subset\mathbb{R}^{d}, ν=α−d/2>0\nu= \alpha- d/2 > 0 describes the smoothness, ρ=2νκ>0\rho= \frac{\sqrt{2\nu}}{\kappa} > 0 for the correlation length, Δ\Delta is the Laplacian operator, and W(x)W(x) is a spatial Gaussian white noise.

The Matérn field provides a canonical example of Gaussian measures. On Rd\mathbb{R}^{d}, or under periodic boundary conditions, the spectral solution of uu defined by Eq. (1) is given by

u(x)=F−1[G(k)W^(k)](x),G(k)=(κ2+∥k∥2)−α/2,(2)u(x) = \mathcal{F}^{-1}\left[G(k)\widehat{W}(k)\right](x), \qquad G(k) = \left(\kappa^{2} + \lVert k\rVert^{2}\right)^{-\alpha/2}, \tag*{(2)}

where F\mathcal{F} is the Fourier transform defined in Appendix C.12, and G(k)G(k) is the Fourier multiplier of the solution operator. The corresponding spectral density of uu is Su(k)∝∣G(k)∣2=(κ2+∥k∥2)−αS_{u}(k) \propto\lvert G(k)\rvert^{2} = \left(\kappa^{2} + \lVert k\rVert^{2}\right)^{-\alpha}. The covariance function of uu is then given by:

c(x)=σ22ν−1Γ(ν)(κ∥x∥)νKν(κ∥x∥),(3)c(x) = \frac{\sigma^{2}}{2^{\nu-1}\Gamma(\nu)}(\kappa\lVert x\rVert)^{\nu}K_{\nu}(\kappa\lVert x\rVert), \tag*{(3)}

where ν=α−d/2\nu= \alpha- d/2, ρ=2νκ\rho= \frac{\sqrt{2\nu}}{\kappa}, σ2=Γ(ν)(4π)d/2κ2νΓ(α)\sigma^{2} = \frac{\Gamma(\nu)}{(4\pi)^{d/2}\kappa^{2\nu}\Gamma(\alpha)} and KνK_{\nu} is the modified Bessel function of the second kind. Detailed derivations are provided in Appendix C.2.3. In addition, Gaussian noise, white noise, and general Matérn fields are closely related, as summarized in Table 1. Here, “Normalized” indicates whether the corresponding noise has been rescaled to have unit marginal variance in the discretized setting. In the remainder of this paper, we use the term Matérn noise to denote noise generated from a Matérn field.

MatérnGaussianWhite
Smoothness α\alpha>d/2> d/20000
Range parameter κ\kappa>0> 0––
Normalized–YesNo

Table 1. Comparison of different random fields.

Having introduced random fields and the Matérn field, we now turn to the SPDEs of interest in this work. In general, SPDEs can be divided into two categories [42]: stationary SPDEs, which describe mappings between probability measures as introduced in Appendix § 7.§ 3, and semilinear evolution equations, which describe the temporal evolution of systems perturbed by noise. Analogous to ordinary time-dependent PDEs, semilinear evolution SPDEs are treated as semilinear stochastic ordinary differential equations (SODEs) on a Hilbert space HU\mathcal{H}_U, taking the form:

{du=[Lu+F(u)]dt+G(u)dW(t),u(0,x)=u0,x∈D.(4)\begin{cases} \mathrm{d}u = [\mathcal{L}u + F(u)]\mathrm{d}t + G(u)\mathrm{d}W(t),\\ u(0,x) = u_0,\quad x \in D. \tag*{(4)} \end{cases}

Here, DD is a bounded spatial domain in Rd\mathbb{R}^d, L\mathcal{L} denotes a linear operator that generates the semigroup Pt=etLP_t = e^{t\mathcal{L}}; W(t)W(t) is an Hξ\mathcal{H}_{\xi}-valued QQ-Wiener process, where Hξ\mathcal{H}_{\xi} is a separable Hilbert space; F:HU→HUF:\mathcal{H}_U \to\mathcal{H}_U denotes a nonlinear operator; and G:HU→L(H0,ξ,HU)G:\mathcal{H}_U \to\mathcal{L}(\mathcal{H}_{0,\xi},\mathcal{H}_U) is a Hilbert–Schmidt operator associated with the Cameron–Martin space H0,ξ\mathcal{H}_{0,\xi} of Hξ\mathcal{H}_{\xi}. Under appropriate dissipativity and ergodicity conditions (see Thm. C.19), the SPDE admits a unique invariant measure μ\mu, which is defined as follows:

Definition 2.3 (Invariant Measure). A Borel probability measure μ\mu on (H,B(H))(H,\mathcal{B}(H)) is called an invariant measure for Equ. (4) if it is invariant under the semigroup (Pt)t≥0(P_t)_{t\geq0}, i.e. ∀ t≥0\forall\,t\geq0, ∀ φ∈Bb(H)\forall\,\varphi\in B_b(H),

∫H(Ptφ)(x) μ(dx)=∫Hφ(x) μ(dx),\int_H (P_t\varphi)(x)\,\mu(\mathrm{d}x) = \int_H \varphi(x)\,\mu(\mathrm{d}x),

Equivalently, if u(0)∼μu(0)\sim\mu, then u(t)∼μu(t)\sim\mu for every $t\geq 0.

Moreover, we say the SPDE is ergodic (or mixing) with invariant measure μ\mu if for every initial condition u0∈Hu_0\in H, Law⁡(u(t))→μ\operatorname{Law}(u(t))\to\mu as t→∞t\to\infty, where →\to denotes weak convergence of probability measures on HH. Further details on the existence and properties of invariant measures for SPDEs are provided in Appendix § 7.§ 3.

Infinite-dimensional Flow Matching

Compared with finite-dimensional generative models, the marginal velocity field in infinite-dimensional generative modeling is well-defined only under rather stringent theoretical conditions [48, 29]. Here, we briefly review the theoretical framework for infinite-dimensional generative models and the conditions required for well-posedness.

Assume that the initial noise g∈Ug\in U follows a Gaussian measure μ0\mu_0 with covariance operator Q:U→UQ:U\to U, where QQ is self-adjoint, nonnegative, and trace-class, and (U,⟨⋅,⋅⟩)(U,\langle\cdot,\cdot\rangle) is a separable Hilbert space. Data samples are denoted by ff, distributed according to the target measure μ1\mu_1. Our objective is to construct an infinite-dimensional flow ψt:[0,1]×U→U\psi_t:[0,1]\times U\to U that pushes μ0\mu_0 to μ1\mu_1. Then the marginal probability measure at time tt is defined as μt:=[ψt]#μ0\mu_t := [\psi_t]_{\#}\mu_0, which admits the representation μt(A)=∫Uμt(A∣f) dμ1(f)\mu_t(A)=\int_U\mu_t(A\mid f)\,\mathrm{d}\mu_1(f), ∀A∈B(U)\forall A\in\mathcal{B}(U), where μt(⋅∣f)\mu_t(\cdot\mid f) denotes the conditional measure and B(U)\mathcal{B}(U) is the Borel σ\sigma-algebra on UU. Then we define the map ψt\psi_t which is associated with a time-dependent velocity field vt:[0,1]×U→Uv_t:[0,1]\times U\to U through

ddtψt(g)=vt(ψt(g)),ψ0(g)=g∼μ0.(5)\frac{\mathrm{d}}{\mathrm{d}t}\psi_t(g)=v_t(\psi_t(g)),\qquad\psi_0(g)=g\sim\mu_0. \tag*{(5)}

When the target measure μ1\mu_1 is fixed, the central object of interest is the marginal velocity field vtv_t, whose regularity and well-posedness directly determine the stability and learnability of the generative flow. The corresponding infinite-dimensional flow matching objective is LFFM=Et,h∼μt[∥vt(h)−vθ(h)∥2]\mathcal{L}_{\mathrm{FFM}}=\mathbb{E}_{t,h\sim\mu_t}\left[\|v_t(h)-v_\theta(h)\|^2\right]. As in finite-dimensional settings, the marginal velocity field is generally intractable. We therefore consider conditional velocity fields vt(⋅∣f)v_t(\cdot\mid f) associated with conditional measures μt(⋅∣f)\mu_t(\cdot\mid f). We recall two standard results from functional flow matching that express the marginal field through conditional dynamics and characterize the required absolute continuity; our new source-design analysis begins in § 3.2.

References

  1. [1]Michael Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. Journal of Machine Learning Research, 26(209):1–80, 2025.
  2. [2]Michael Samuel Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In The Eleventh International Conference on Learning Representations, 2023.arxiv.org/abs/2209.15571
  3. [3]Alain Bensoussan. Stochastic navier-stokes equations. Acta Applicandae Mathematica, 38(3):267–304, 1995.DOI
  4. [4]Nicholas M. Boffi, Michael S. Albergo, and Eric Vanden-Eijnden. Flow map matching with stochastic interpolants: A mathematical framework for consistency models, 2025. URL https://arxiv.org/abs/2406.07507.
  5. [5]René Carmona. Stochastic Partial Differential Equations: Six Perspectives: Six Perspectives. Number 64. American Mathematical Soc., 1999.
  6. [6]Ting Chen. On the importance of noise scheduling for diffusion models, 2023. URL https://arxiv.org/abs/2301.10972.
  7. [8]Yifan Chen and Eric Vanden-Eijnden. Scale-adaptive generative flows for multiscale scientific data. arXiv preprint arXiv:2509.02971, 2025.
  8. [9]Yifan Chen, Eric Vanden-Eijnden, and Jiawei Xu. Lipschitz-guided design of interpolation schedules in generative models, 2025. URL https://arxiv.org/abs/2509.01629.
  9. [10]Yuzhu Chen, Fengxiang He, Shi Fu, Xinmei Tian, and Dacheng Tao. Adaptive time-stepping schedules for diffusion models. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024.
  10. [11]Giuseppe Da Prato and Jerzy Zabczyk. Ergodicity for infinite dimensional systems, volume 229. Cambridge university press, 1996.
  11. [12]Giuseppe Da Prato and Jerzy Zabczyk. Stochastic Equations in Infinite Dimensions. Encyclopedia of Mathematics and its Applications. Cambridge University Press, 2 edition, 2014.
  12. [13]J-P Eckmann and M Hairer. Invariant measures for stochastic partial differential equations in unbounded domains. Nonlinearity, 14(1):133–151, November 2000. ISSN 1361-6544. doi: 10.1088/0951-7715/14/1/308. URL http://dx.doi.org/10.1088/0951-7715/14/1/308.
  13. [14]Jiajie Fan, Laure Vuaille, Thomas Bäck, and Hao Wang. On the noise scheduling for generating plausible designs with diffusion models. arXiv preprint arXiv:2311.11207, 2023.
  14. [15]Giulio Franzese, Giulio Corallo, Simone Rossi, Markus Heinonen, Maurizio Filippone, and Pietro Michiardi. Continuous-time functional diffusion processes. Advances in Neural Information Processing Systems, 36:37370–37400, 2023.arxiv.org/abs/2303.00800
  15. [16]Shiqi Gong, Peiyuan Hu, Qi Meng, Yue Wang, Rongchan Zhu, Bingguang Chen, Zhiming Ma, Hao Ni, and Tie-Yan Liu. Deep latent regularity network for modeling stochastic partial differential equations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 7740–7747, 2023.
  16. [17]Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014.
  17. [18]Qiushan Guo, Sifei Liu, Yizhou Yu, and Ping Luo. Rethinking the noise schedule of diffusion-based generative models. 2023.
  18. [19]Martin Hairer. Solving the kpz equation. Annals of mathematics, pages 559–664, 2013.
  19. [20]Martin Hairer and Jonathan C Mattingly. Ergodicity of the 2d navier-stokes equations with degenerate stochastic forcing. Annals of Mathematics, pages 993–1032, 2006.
  20. [21]Tiankai Hang, Shuyang Gu, Jianmin Bao, Fangyun Wei, Dong Chen, Xin Geng, and Baining Guo. Improved noise schedule for diffusion training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4796–4806, 2025.arxiv.org/abs/2407.03297
  21. [22]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020.arxiv.org/abs/2006.11239
  22. [23]Ulrich Hornung. Applications of the homogenization method to flow and transport in porous media. Summer School on Flow and Transport in Porous Media, pages 167–222, 1992.DOI
  23. [24]Peiyan Hu, Qi Meng, Bingguang Chen, Shiqi Gong, Yue Wang, Wei Chen, Rongchan Zhu, Zhi-Ming Ma, and Tie-Yan Liu. Neural operator with regularity structure for modeling dynamics driven by spdes, 2022. URL https://arxiv.org/abs/2204.06255.
  24. [25]Peiyan Hu, Haodong Feng, Yue Wang, and Zhiming Ma. Model-based closed-loop control algorithm for stochastic partial differential equation control, 2025. URL https://arxiv.org/abs/2505.05521.
  25. [26]Jiahe Huang, Guandao Yang, Zichen Wang, and Jeong Joon Park. Diffusionpde: Generative pde-solving under partial observation. Advances in Neural Information Processing Systems, 37:130291–130323, 2024.
  26. [27]Kerson Huang. Introduction to statistical physics. Chapman and Hall/CRC, 2009.
  27. [28]Gavin Kerrigan, Justin Ley, and Padraic Smyth. Diffusion generative models in infinite dimensions. arXiv preprint arXiv:2212.00886, 2022.arxiv.org/abs/2212.00886
  28. [29]Gavin Kerrigan, Giosue Migliorini, and Padraic Smyth. Functional flow matching. arXiv preprint arXiv:2305.17209, 2023.arxiv.org/abs/2305.17209
  29. [30]S Kida and Y Murakami. Kolmogorov similarity in freely decaying turbulence. The Physics of fluids, 30(7):2030–2039, 1987.DOI
  30. [31]Junwan Kim, Jiho Park, Seonghu Jeon, and Seungryong Kim. Better source, better flow: Learning condition-dependent source distribution for flow matching. arXiv preprint arXiv:2602.05951, 2026. doi: 10.48550/arXiv.2602.05951.DOI
  31. [32]Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.arxiv.org/abs/1312.6114
  32. [33]Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Aziz zadensheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1–97, 2023.
  33. [34]Zhiqi Li, Yuchen Sun, Greg Turk, and Bo Zhu. Functional mean flow in hilbert space. arXiv preprint arXiv:2511.12898, 2025. doi: 10.48550/arXiv.2511.12898.DOI
  34. [35]Zongyi Li, Nikola Kovachki, Kamyar Aziz zadensheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations, 2021. URL https://arxiv.org/abs/2010.08895.
  35. [36]Jae Hyun Lim, Nikola B Kovachki, Ricardo Baptista, Christopher Beckham, Kamyar Aziz zadensheli, Jean Kossaifi, Vikram Voleti, Jiaming Song, Karsten Kreis, Jan Kautz, et al. Score-based diffusion models in function space. arXiv preprint arXiv:2302.07400, 2023.
  36. [37]Finn Lindgren, Håvard Rue, and Johan Lindström. An explicit link between gaussian fields and gaussian markov random fields: the stochastic partial differential equation approach. Journal of the Royal Statistical Society Series B: Statistical Methodology, 73(4):423–498, 2011.
  37. [39]Pengwei Liu, Qinxin Wu, Xingyu Ren, Yian Wang, and Dong Ni. A deep-learning-based surrogate modeling method with application to plasma processing. Chemical Engineering Research and Design, 211:299–317, 2024b.DOI
  38. [41]Pengwei Liu, Pengkai Wang, Xingyu Ren, Hangjie Yuan, Zhongkai Hao, Chao Xu, Shengze Cai, and Dong Ni. Aerogto: An efficient graph-transformer operator for learning large-scale aerodynamics of 3d vehicle geometries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18924–18932, 2025b.DOI
  39. [42]Gabriel J. Lord, Catherine E. Powell, and Tony Shardlow. An Introduction to Computational Stochastic PDEs. Cambridge Texts in Applied Mathematics. Cambridge University Press, 2014.DOI
  40. [43]Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research, pages 1–22, 2025.DOI
  41. [45]Shimon Malnick, Matan Rusanovsky, Ohad Fried, and Shai Avidan. Optimal transport flow matching by design. arXiv preprint arXiv:2606.04092, 2026. doi: 10.48550/arXiv.2606.04092.DOI
  42. [46]Kunwoo Na, Junghyun Lee, Se-Young Yun, and Sungbin Lim. Probability-flow ode in infinite-dimensional function spaces. arXiv preprint arXiv:2503.10219, 2025.arxiv.org/abs/2503.10219
  43. [47]Ariel Neufeld and Philipp Schmock­er. Solving stochastic partial differential equations using neural networks in the wiener chaos expansion. arXiv preprint arXiv:2411.03384, 2024.arxiv.org/abs/2411.03384
  44. [48]Jakiw Pidstrigach, Youssef Marzouk, Sebastian Reich, and Sven Wang. Infinite-dimensional diffusion models. Journal of Machine Learning Research, 25(414):1–52, 2024.arxiv.org/abs/2302.10130
  45. [49]Md Ashiqur Rahman, Manuel A Florez, Anima Anandkumar, Zachary E Ross, and Kamyar Aziz-zadenesheli. Generative adversarial neural operators. arXiv preprint arXiv:2205.03017, 2022.arxiv.org/abs/2205.03017
  46. [50]Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378:686–707, 2019.DOI
  47. [51]Christopher Salvi, Maud Lemercier, and Andris Gerasimovics. Neural stochastic pdes: Resolution-invariant learning of continuous spatiotemporal dynamics. Advances in Neural Information Processing Systems, 35:1333–1344, 2022.arxiv.org/abs/2110.10249
  48. [52]Simo Särkkä and Arno Solin. Applied stochastic differential equations, volume 10. Cambridge University Press, 2019.DOI
  49. [53]Jacob H Seidman, Georgios Kissas, George J Pappas, and Paris Perdikaris. Variational autoencoding neural operators. arXiv preprint arXiv:2302.10351, 2023.arxiv.org/abs/2302.10351
  50. [54]Yuyang Shi, Valentin De Bortoli, Andrew Campbell, and Arnaud Doucet. Diffusion schrödinger bridge matching. Advances in Neural Information Processing Systems, 36:62183–62223, 2023.arxiv.org/abs/2303.16852
  51. [56]Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021.
  52. [57]Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research, pages 1–34, 2024.arxiv.org/abs/2302.00482
  53. [58]Weijun Xu, Wenhao Zhao, and Shuhan Zhou. Interface fluctuations for 1d stochastic allen-cahn equation revisited: W. xu et al. Archive for Rational Mechanics and Analysis, 249(5):52, 2025.arxiv.org/abs/2402.19070
  54. [59]Jiachen Yao, Abbas Mammadov, Julius Berner, Gavin Kerrigan, Jong Chul Ye, Kamyar Azizzadenesheli, and Anima Anandkumar. Guided diffusion sampling on function spaces with applications to pdes. arXiv preprint arXiv:2505.17004, 2025. doi: 10.48550/arXiv.2505.17004.arxiv.org/abs/2505.17004
  55. [60]Dongkun Zhang, Ling Guo, and George Em Karniadakis. Learning in modal space: Solving time-dependent stochastic pdes using physics-informed neural networks. SIAM Journal on Scientific Computing, 42(2):A639–A665, 2020.arxiv.org/abs/1905.01205
  56. [61]Pascal Zwick, Nils Friederich, Maximilian Beichter, Lennart Hilbert, Ralf Mikut, and Oliver Bringmann. Lediflow: Learned distribution-guided flow matching to accelerate image generation. arXiv preprint arXiv:2505.20723, 2025. doi: 10.48550/arXiv.2505.20723.DOI

Paper details

vlnc:W_ekg8t3gu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 16

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents