betaFeedback

WorksComputer Science

WaSP: Wasserstein Data Synthesis Pipeline for Solver-Free Operator Learning

Abstract

Neural operators offer fast surrogates for solving PDEs, but training them typically requires expensive input-solution pairs. Two solver-free alternatives include physics-informed learning and training on manufactured data. Physics-informed learning approximates solutions by minimizing PDE residuals. While the residual norm must be strong enough to control solution error, stronger norms increase mini-batch variance and hinder optimization. Manufactured data constructs input-solution pairs for learning by applying the PDE operator to artificial solutions. However, the resulting training distribution can be mutually singular with the target distribution, yielding poor accuracy on the target distribution. We introduce the Wasserstein Data Synthesis Pipeline (WaSP), which trains on manufactured input-solution pairs while gradually mitigating the distribution shift. WaSP synthesizes new pairs by applying the learned operator to a path of intermediate distributions converging to the target input distribution. Training proceeds in stages along this path. At each stage, the previous surrogate generates solutions, and the exact PDE operator derives corresponding inputs, yielding training pairs for the next surrogate. Theoretically, we prove a stepwise error-propagation bound for the distribution shift introduced during synthesis. Empirically, WaSP reduces solution error by 3.9×3.9\times to 9.5×9.5\times compared to existing solver-free methods on PDEs such as Allen-Cahn, Kuramoto-Sivashinsky, and Ginzburg-Landau, and better reproduces the two chaotic systems' statistical observables.

Introduction

Operator learning constructs mappings between function spaces, such as solution operators that map input conditions to the corresponding solutions of partial differential equations (PDEs). Neural operators parametrize these maps and, once trained, replace numerical solvers at inference time (Li

©2026 Ryan Liu, Chuwei Wang, Ricardo Baptista, Andrew Stuart and Anima Anandkumar.

License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. et al., 2021; Lu et al., 2021; Kovachki et al., 2023). They are now applied where repeated high-fidelity simulation is expensive or infeasible, including fluid dynamics [10], weather forecasting [43], quantum many-body dynamics (Waugh et al., 2026), astrophysical accretion and feedback [6], and chaotic-system surrogates [53].

However, the standard training pipeline remains data-intensive. Mainstream methods are trained by supervised regression on input–solution pairs, often obtained from numerical solvers, laboratory measurements, or observational data. Acquiring them at scale can be challenging; simulating a chaotic or multiscale system often requires fine spatial–temporal resolution and expensive implicit solves, and laboratory and observational data are sparse, noisy, and constrained by the measurement setup. Large curated datasets are available for selected benchmark and operational settings (Li et al., 2008; Hersbach et al., 2020; Takamoto et al., 2022; Ohana et al., 2024), but not for all equations, parameter regimes, geometries, or physical systems. We therefore study solver-free operator learning, that is, training a neural operator without any solver-generated solution label for the target inputs.

Physics-informed learning is a mainstream solver-free operator learning paradigm, which replaces solution labels with an equation-derived loss. Physics-informed neural networks (PINNs) instantiate the idea for learning a single PDE instance; it directly parametrizes the solution function with a neural network that maps points in its domain to the corresponding solution values and learns from the PDE residual and boundary constraints [45]. Physics-informed neural operators (PINOs) extend the principle from single PDE instances to operator learning (Li et al., 2024; Goswami et al., 2023). However, training with physics-informed objectives encounters two obstructions. (1) A small residual loss is not automatically a reliable certificate of solution quality. Its validity depends on whether the residual norm aligns with the stability properties of the PDE [52]. (2) Nonlinear terms and high-order derivatives in the PDE operator enter the computation graph directly, and the resulting composite objective is difficult to optimize [55, 30]. We survey the proposed mitigations in § 7.

Another solver-free route reverses the usual data-generation direction. Instead of solving the PDE from inputs to solutions, it samples artificial solutions and applies the PDE operator to obtain the corresponding input conditions[23, 48, 7]. Applying the PDE operator is far cheaper and more parallelizable than a numerical solve, so we can efficiently generate large volumes of labeled input–solution pairs. Given the resemblance of this approach to the method of manufactured solutions used to verify numerical solvers [46], we refer to these generated pairs as manufactured data (MD) in this work. However, since the training pairs are sampled and constructed from the artificial solution prior, their training distribution diverges from the target distribution and provides limited learning signal that is relevant for improving test-time performance.2

Our Contribution. We analyze both routes and develop an algorithm that avoids their obstructions. For physics-informed neural operators, we identify a regularity–variance tradeoff (§ 3.1 and § 3.2). On the Kuramoto–Sivashinsky equation, we construct a sequence of solutions whose residual and initial-condition errors both vanish in a weak norm while the solution error does not, and we show that even when the norm is strong enough to admit a stability estimate, it carries an amplification constant growing exponentially in the inverse dissipation. On the other hand, while stronger norms restore control, we show with a linear model that such norms amplify the mini-batch variance of the residual learning objective, posing significant challenges for optimizing stability-compatible norms in practice. For manufactured data methods, we prove a complementary obstruction (§ 3.3). In Gaussian function-space models, the induced input distribution is mutually singular with the target input distribution unless its variance tracks the target’s mode by mode. Without further assumptions on the neural operator used, the mutual singularity implies that learning on the exact manufactured pairs provides no performance guarantees at the target distribution. Calibrating the artificial solution prior mode-by-mode to avoid the obstruction, we argue, requires distributional knowledge that is itself as hard as solving the equation.

This diagnosis motivates the Wasserstein Data Synthesis Pipeline (WaSP), a solver-free training method that preserves the optimization advantage of manufactured-data supervision while mitigating its distribution shift by training along a transport path (§ 4). WaSP samples artificial solutions and applies the PDE operator to obtain the manufactured input distribution, estimates a transport map that maps the manufactured inputs to the target input distribution, and applies McCann interpolation to define a sequence of intermediate training distributions ([39, 51]). Training thus becomes a supervised curriculum ordered from easy to hard ([4]). At each stage, the current operator synthesizes solutions for inputs drawn from the next intermediate distribution, expecting that the neural operator generalizes locally and synthesizes well-approximated solutions. WaSP applies the PDE operator to the solutions and returns corresponding inputs, forming exact input–solution pairs for training at the next stage. The primary learning signal is the solution-space regression loss on the pairs, which avoids the optimization pathologies of physics-informed methods. WaSP breaks the large structural distribution shift between manufactured data and target down into small and manageable local generalization steps; therefore, it also avoids the obstruction identified for manufactured data. Figure 1 visualizes the transport path for the Kuramoto–Sivashinsky equation, showing how it gradually bridges the distribution shift by defining the intermediate distributions.

Figure 1. Illustration of how WaSP training distributions are constructed. Top: the McCann interpolation computed for 2D Gaussian point clouds. The gray segments are the estimated optimal transport map TT between samples drawn from a reference distribution ν0\nu_0 and the target ν1\nu_1. Each intermediate distribution is defined by the sampling rule in (17), where a sample from the reference is drawn, mapped by TT, and linearly interpolated. Bottom: an input–solution pair for the Kuramoto–Sivashinsky equation interpolated between the manufactured input distribution (s=0s = 0, left) and the target input distribution (s=1s = 1, right). The top row is the forcing component of the input, which anneals to near zero as ss increases; the bottom row is the corresponding solution field u(t,x)u(t, x), obtained from an ETD-RK4 solver. As the forcing decreases in magnitude and the initial state moves onto the target attractor, the chaotic dynamics emerge, so the interpolation presents the operator with a sequence of progressively harder problems, forming a training curriculum.

We deliver three primary results to support WaSP. First, theoretically, we prove a stepwise error-propagation bound under strong Lipschitzness assumptions (Theorem 12). We illustrate how, without the local generalization assumptions, the error propagation bound worsens when the transport path is refined, and with local generalization, the error propagation bound can be exponentially improved and tightens instead. Second, implementation-wise, we argue in § 4.2 how, in practice, the strong and unrealistic Lipschitzness assumptions can be removed, which motivates two important designs, the local PDE residual regularization and the adaptive scheduling function. We also discuss efficient implementation of WaSP and methods to improve its numerical stability in § 4.3. Finally, empirically, with no solver-generated labeled pair from the target distribution used in training, WaSP reduces relative solution error by factors of 3.9, 9.5, and 4.6 relative to the strongest baseline on the Allen–Cahn, Kuramoto–Sivashinsky, and complex Ginzburg–Landau equations. On the two chaotic systems, it also tracks statistical observables more closely, including the spatial and temporal autocorrelations, energy spectrum, and invariant measures. We also conduct two ablations: one removes the self-generated targets while keeping the transport curriculum, and the other adds physics-informed objectives to the manufactured-data training. The ablation results are consistent with the obstructions identified in § 3.

Paper Organization. Section 2 defines the PDE and operator-learning notation, collects the regularity and stability tools used in the analysis, and states the training objectives we compare. Section 3 formalizes the two obstructions: the regularity–variance tradeoff for residual training (§ 3.1 and § 3.2) and the structural distribution shift for manufactured data (§ 3.3). Section 4 introduces WaSP, proves the stepwise error-propagation bound, and shows how the local residual regularizer and the adaptive schedule relax its unrealistic assumptions (§ 4.2), and states the algorithm with its implementation (§ 4.3). Section 5 describes the three equations, the experiment and evaluation setup, and reports pointwise errors and long-time observables against the physics-informed and manufactured-data baselines. Section 6 discusses construction of the manufactured solution distribution (§ 6.1) and the cost of the transport estimate (§ 6.2). Section 7 places WaSP in context, and Section 8 concludes. Proofs and hyperparameters are in the appendix.

Preliminaries

his section first introduces the PDE and operator-learning setting (§ 2.2), then defines the regularity and stability notions used in our analysis (§ 2.3), and finally formalizes the training objectives considered in this work (§ 2.4).

PDE Setting and Notation

We consider a general PDE setting in which the unknown solution field uu is related to the prescribed input vv through an equation

N(u)=v.\mathcal{N}(u)=v.

Here N:U→V\mathcal{N}:\mathcal{U}\to\mathcal{V} denotes the PDE operator written in a unified form, U\mathcal{U} is a normed space of solution fields, and V\mathcal{V} is a normed space of PDE inputs, which may include forcing terms, boundary data, initial data, or tuples of these quantities. Throughout, we assume U\mathcal{U} and V\mathcal{V} are chosen so that N\mathcal{N} is a bijection whose inverse is the solution operator G:V→U\mathcal{G}:\mathcal{V}\to\mathcal{U}, giving G∘N=Id⁡\mathcal{G}\circ\mathcal{N}=\operatorname{Id} on U\mathcal{U} and N∘G=Id⁡\mathcal{N}\circ\mathcal{G}=\operatorname{Id} on V\mathcal{V}; this is an assumption of existence and uniqueness only. Continuous dependence of uu on vv is a separate property, formalized in Definition 2, and it is precisely the property that degrades for the equations we study. Section 5.1 describes how including the forcing field in the input realizes this assumption for the three benchmarks.

Example 1 (Evolution problem). Consider an evolution problem on the space–time domain [0,T]×Ω[0,T]\times\Omega, where Ω⊂Rd\Omega\subset\mathbb{R}^{d} is the spatial domain:

{Dt,x(u)(t,x)=f(t,x),(t,x)∈(0,T]×Ω,I(u)(x)=u0(x),x∈Ω,B(u)(t,x)=g(t,x),(t,x)∈(0,T]×∂Ω.\begin{cases} \mathcal{D}_{t,x}(u)(t,x)=f(t,x), & (t,x)\in(0,T]\times\Omega,\\ \mathcal{I}(u)(x)=u_{0}(x), & x\in\Omega,\\ \mathcal{B}(u)(t,x)=g(t,x), & (t,x)\in(0,T]\times\partial\Omega. \end{cases}

where Dt,x\mathcal{D}_{t,x} is the space–time PDE operator, I\mathcal{I} is the initial-condition operator, B\mathcal{B} is the boundary operator, and all three may be nonlinear. In the unified notation,

N(u)=(I(u),Dt,x(u),B(u)),v=(u0,f,g).\mathcal{N}(u)=\bigl(\mathcal{I}(u),\mathcal{D}_{t,x}(u),\mathcal{B}(u)\bigr),\qquad v=(u_{0},f,g).

so V\mathcal{V} is the product of the initial, interior, and boundary condition spaces. The benchmarks of Section 5.1 are of this form with periodic boundaries, giving v=(u0,f)v=(u_{0},f).

Operator Learning

Operator learning seeks a parametric approximation Gθ:V→U\mathcal{G}_{\theta}:\mathcal{V}\to\mathcal{U} of the solution operator G\mathcal{G}. In realistic applications, the inputs of interest are governed by an underlying distribution over forcing fields, ensembles of initial conditions, or observational states encountered in deployment. We denote this target input distribution by ν1∈P(V)\nu_{1}\in\mathcal{P}(\mathcal{V}), where for a normed space Z\mathcal{Z} we write P(Z)\mathcal{P}(\mathcal{Z}) for the set of Borel probability distributions on Z\mathcal{Z} with the moments required by the Wasserstein distances used below. Unless otherwise specified, ν1\nu_{1} denotes the target input distribution and ν\nu denotes a generic input distribution. The subscript anticipates Section 4, where ν1\nu_{1} is the endpoint of a transport path that starts at a chosen reference distribution ν0\nu_{0}, such as the input distribution induced by manufactured solutions. For a normed evaluation space X⊆U\mathcal{X}\subseteq\mathcal{U}, the target risk of Gθ\mathcal{G}_{\theta} under a generic input distribution ν\nu is

E(θ;ν;X):=Ev∼ν[∥Gθ(v)−G(v)∥X2].(1)\mathcal{E}(\theta;\nu;\mathcal{X}) := \mathbb{E}_{v\sim\nu}\left[\left\|\mathcal{G}_{\theta}(v)-\mathcal{G}(v)\right\|_{\mathcal{X}}^{2}\right]. \tag*{(1)}

Thus, the definition applies to any ν\nu, and evaluating it at ν=ν1\nu=\nu_{1} gives the target-distribution risk.

A neural operator parametrizes Gθ\mathcal{G}_{\theta} by composing pointwise lifting and projection maps with a sequence of layers that combine a local linear map with a learnable integral operator on the physical domain [29]. We use the Fourier Neural Operator (Li et al., 2021) for Gθ\mathcal{G}_{\theta} in all of our experiments, which parametrizes the integral operator as a spectral convolution operator. The training procedures developed in this work are independent of the architectural choices.

Regularity and Stability

Our analysis compares losses and errors measured in norms of different regularity. We therefore recall the relevant Sobolev spaces and formalize what it means for one norm to be stronger than another.

Let Ω⊂Rd\Omega\subset\mathbb{R}^{d} be a domain and let m∈N0m\in\mathbb{N}_{0}. We write Hm(Ω):=Wm,2(Ω)H^{m}(\Omega):=W^{m,2}(\Omega) for the Sobolev space of functions whose weak derivatives up to order mm belong to L2(Ω)L^{2}(\Omega). For a real index s∈Rs\in\mathbb{R}, the periodic Sobolev space Hs(Td)H^s(\mathbb{T}^d) is defined by

∥f∥Hs(Td)2:=∑k∈Zd(1+∣k∣2)s∣f^(k)∣2,(2)\lVert f\rVert_{H^s(\mathbb{T}^d)}^2 := \sum_{k\in\mathbb{Z}^d} (1+|k|^2)^s |\widehat{f}(k)|^2, \tag*{(2)}

where f^(k)\widehat{f}(k) is the kth Fourier coefficient of ff.

Definition 2 (PDE stability). Let A⊆U\mathcal{A} \subseteq\mathcal{U} be an admissible set, let X\mathcal{X} be an evaluation space with A⊆X\mathcal{A} \subseteq\mathcal{X}, and let Y\mathcal{Y} be a residual space with N(A)⊆Y\mathcal{N}(\mathcal{A}) \subseteq\mathcal{Y}. The solution operator G\mathcal{G} is stable from Y\mathcal{Y} to X\mathcal{X} on A\mathcal{A} if there exists a constant C>0C > 0 such that

∥u−u~∥X≤C∥N(u)−N(u~)∥Yfor all u,u~∈A.(3)\lVert u-\widetilde{u}\rVert_{\mathcal{X}} \le C\lVert\mathcal{N}(u)-\mathcal{N}(\widetilde{u})\rVert_{\mathcal{Y}} \qquad\text{for all } u,\widetilde{u}\in\mathcal{A}. \tag*{(3)}

Equivalently, the restriction of G\mathcal{G} to N(A)\mathcal{N}(\mathcal{A}) is CC-Lipschitz from ∥⋅∥Y\lVert\cdot\rVert_{\mathcal{Y}} to ∥⋅∥X\lVert\cdot\rVert_{\mathcal{X}}.

Stability converts a bound on the PDE residual into a bound on the solution error. If v∈Vv \in\mathcal{V} satisfies G(v)∈A\mathcal{G}(v) \in\mathcal{A}, then taking u=G(v)u = \mathcal{G}(v) in (3) and using N∘G=Id⁡\mathcal{N}\circ\mathcal{G}=\operatorname{Id} gives

∥G(v)−u~∥X≤C∥N(u~)−v∥Yfor every trial solution u~∈A.(4)\lVert\mathcal{G}(v)-\widetilde{u}\rVert_{\mathcal{X}} \le C\lVert\mathcal{N}(\widetilde{u})-v\rVert_{\mathcal{Y}} \qquad\text{for every trial solution } \widetilde{u}\in\mathcal{A}. \tag*{(4)}

Thus, the residual measured in Y\mathcal{Y} certifies the solution error measured in X\mathcal{X}, with a factor of CC. Whether such a CC exists depends on both norms, which motivates comparing them.

Definition 3 (Norm comparison). Let ∥⋅∥a\lVert\cdot\rVert_a and ∥⋅∥b\lVert\cdot\rVert_b be norms on a common vector space Z\mathcal{Z}. We say that ∥⋅∥b\lVert\cdot\rVert_b is stronger than ∥⋅∥a\lVert\cdot\rVert_a on Z\mathcal{Z} if there exists a constant C>0C > 0 such that

∥f∥a≤C∥f∥bfor every f∈Z,(5)\lVert f\rVert_a \le C\lVert f\rVert_b \qquad\text{for every } f\in\mathcal{Z}, \tag*{(5)}

and strictly stronger if no converse inequality holds uniformly on Z\mathcal{Z}.

On the Sobolev scale, Hs2H^{s_2} is stronger than Hs1H^{s_1} when s2≥s1s_2 \ge s_1, so a stronger norm controls more derivatives and imposes higher regularity. Combining (5) and (3) show that stability is preserved when the residual norm is strengthened or the solution norm is weakened and may fail in either of the opposite directions.

Training Neural Operators

We now introduce the three training objectives considered in this work. We call a training procedure solver-free if it never uses a solver-generated pair (v,G(v))(v,\mathcal{G}(v)) as a training label; it may sample manufactured solutions u∼μu \sim\mu from a distribution μ∈P(U)\mu\in\mathcal{P}(\mathcal{U}), evaluate the PDE operator N\mathcal{N}, and draw unpaired inputs v∼ν1v \sim\nu_1 from the target distribution. The distribution μ\mu may itself be produced by any procedure other than running a solver, which is detailed in § 6.1.

Data Loss. The supervised data loss minimizes the expected squared error between the predicted and the true solution under an input distribution ν\nu. Writing ℓdata(w;v;X):=∥w−G(v)∥X2\ell_{\mathrm{data}}(w;v;\mathcal{X}) := \lVert w-\mathcal{G}(v)\rVert_{\mathcal{X}}^2 for the per-sample loss at an input v∈Vv \in\mathcal{V} and a predicted solution w∈Uw \in\mathcal{U}, the operator-level loss evaluated on the target distribution ν1\nu_1 is

Ldata(θ;ν1;X):=Ev∼ν1ℓdata(Gθ(v);v;X)=Ev∼ν1[∥Gθ(v)−G(v)∥X2],(6)\mathcal{L}_{\mathrm{data}}(\theta;\nu_1;\mathcal{X}) := \mathbb{E}_{v\sim\nu_1}\ell_{\mathrm{data}}(\mathcal{G}_{\theta}(v);v;\mathcal{X}) = \mathbb{E}_{v\sim\nu_1}\left[\lVert\mathcal{G}_{\theta}(v)-\mathcal{G}(v)\rVert_{\mathcal{X}}^2\right], \tag*{(6)}

where X\mathcal{X} is the normed space used to quantify the error and G(v)\mathcal{G}(v) is obtained from a numerical solver or from experimental measurements. This objective requires labeled pairs (v,G(v))(v,\mathcal{G}(v)) with v∼ν1v \sim\nu_1, which can be expensive or unavailable for new equations or environments.

PDE Residual Loss. Physics-informed training replaces labeled solution pairs with supervision from the governing equation [45]. Specifically, the per-sample residual loss is

ℓPDE(w;v;Y):=∥N(w)−v∥Y2,(7)\ell_{\mathrm{PDE}}(w;v;\mathcal{Y}) := \lVert\mathcal{N}(w)-v\rVert_{\mathcal{Y}}^2, \tag*{(7)}

where v∈Vv \in\mathcal{V} is the input and w∈Uw \in\mathcal{U} is the predicted solution. In the PINN setting, where the learning target is a single PDE instance (i.e., a fixed v∈Vv \in\mathcal{V} and a neural-network-parametrized ww), the infinite-dimensional norm ∥⋅∥Y2\lVert\cdot\rVert_{\mathcal{Y}}^2 is approximated by a Monte-Carlo estimate of the mean-squared residual at a set of collocation points, corresponding to an L2L^2-type norm over the interior residual and, when present, the initial- and boundary-condition mismatches [12].

In operator learning, physics-informed neural operators extend this idea by applying the residual loss to the predicted solution field Gθ(v)\mathcal{G}_{\theta}(v) and averaging over inputs v∼νv \sim\nu [22] (Li et al. 2024). This gives

References

  1. [1]I. S. Aranson and L. Kramer. The world of the complex Ginzburg-Landau equation. Reviews of Modern Physics, 74(1):99, 2002.DOI
  2. [2]E. Arazo, D. Ortego, P. Albert, N. E. O’Connor, and K. McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International joint conference on neural networks (IJCNN), pages 1–8. IEEE, 2020.arxiv.org/abs/1908.02983
  3. [3]S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan. A theory of learning from different domains. Machine Learning, 79(1):151–175, 2010.
  4. [4]Y. Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning, pages 41–48, 2009.DOI
  5. [5]J. A. L. Benitez, T. Furuya, F. Faucher, A. Kratios, X. Tricoche, and M. V. de Hoop. Out-of-distributional risk bounds for neural operators with applications to the helmholtz equation. Journal of Computational Physics, 513:113168, 2024.
  6. [6]N. Bhojwani, C. Wang, H.-Y. Wang, C. Sun, E. R. Most, and A. Anandkumar. From black hole to galaxy: Neural operator: Framework for accretion and feedback dynamics. arXiv preprint arXiv:2512.01576, 2025.arxiv.org/abs/2512.01576
  7. [7]R. Bischof, M. Piovarci, M. Kraus, S. Mishra, and B. Bickel. Hypino: Multi-physics neural operators via hyperpinns and the method of manufactured solutions. Advances in Neural Information Processing Systems, 38:144798–144831, 2026.arxiv.org/abs/2509.05117
  8. [8]Y. Brenier. Polar factorization and monotone rearrangement of vector-valued functions. Communications on Pure and Applied Mathematics, 44(4):375–417, 1991.DOI
  9. [9]S. Cao. Choose a transformer: Fourier or Galerkin. Advances in Neural Information Processing Systems, 34:24924–24940, 2021.arxiv.org/abs/2105.14995
  10. [10]S. Cao, F. Brarda, R. Li, and Y. Xi. Spectral-refiner: Accurate fine-tuning of spatiotemporal fourier neural operator for turbulent flows. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=MKP1g8wUOP.arxiv.org/abs/2405.17211
  11. [11]B. Charlier, J. Feydy, J. A. Glaunès, F.-D. Collin, and G. Durif. Kernel operations on the GPU, with autodiff, without memory overflows. Journal of Machine Learning Research, 22(74):1–6, 2021. URL http://jmlr.org/papers/v22/20-275.html.
  12. [12]Y. Chen, B. Hosseini, H. Owhadi, and A. M. Stuart. Solving and learning nonlinear pdes with gaussian processes. Journal of Computational Physics, 447:110668, 2021.arxiv.org/abs/2103.12959
  13. [13]S. M. Cox and P. C. Matthews. Exponential time differencing for stiff systems. Journal of Computational Physics, 176(2):430–455, 2002.DOI
  14. [14]M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in Neural Information Processing Systems, 26, 2013.
  15. [15]P. Cvitanović, R. L. Davidchack, and E. Siminos. On the state space geometry of the kuramoto–sivashinsky flow in a periodic domain. SIAM Journal on Applied Dynamical Systems, 9(1):1–33, 2010.DOI
  16. [16]G. Da Prato and J. Zabczyk. Stochastic equations in infinite dimensions. Cambridge University Press, 2 edition, 2014.
  17. [17]M. El Hamri, I. Falih, and Y. Rozenholc. McCann’s interpolation for gradual domain adaptation on the Wasserstein geodesic. In 2025 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2025.DOI
  18. [18]P. Gabriel and C. Marco. Computational optimal transport with applications to data sciences. Foundations and trends® in machine learning, 11(5-6):355–607, 2019.DOI
  19. [19]A. Ganesh­ram, H. Maust, V. Duriuisseaux, Z. Li, Y. Wang, D. Leibovici, O. Bruno, T. Hou, and A. Anandakumar. FC-PINO: High precision physics-informed neural operators via fourier continuation. arXiv preprint arXiv:2211.15960, 2022.arxiv.org/abs/2211.15960
  20. [20]S. Ghadimi and G. Lan. Stochastic first- and zeroth-order methods for nonconvex stochastic programming. SIAM Journal on Optimization, 23(4):2341–2368, 2013.arxiv.org/abs/1309.5549
  21. [21]A. Gonzalez-Sanz, J.-M. Loubes, and J. Niles-Weed. Weak limits of entropy regularized optimal transport; potentials, plans and divergences. arXiv preprint arXiv:2207.07427, 2022.
  22. [22]S. Goswami, A. Bora, Y. Yu, and G. E. Karniadakis. Physics-informed deep neural operator networks. In Machine Learning in Modeling and Simulation: Methods and Applications, pages 219–254. Springer, 2023.arxiv.org/abs/2207.05748
  23. [23]E. Hasani and R. A. Ward. Generating synthetic data for neural operators. arXiv preprint arXiv:2401.02398, 2024.arxiv.org/abs/2401.02398
  24. [25]M. Herde, B. Raonić, T. Rohner, R. Käppeli, R. Molinaro, E. De Bezenac, and S. Mishra. Poseidon: Efficient foundation models for PDEs. Advances in Neural Information Processing Systems, 37:72525–72624, 2024.arxiv.org/abs/2405.19101
  25. [27]J. Jiang and Y.-C. Lai. Model-free prediction of spatiotemporal dynamical systems with recurrent neural networks: Role of network spectral radius. Physical Review Research, 1(3):033056, 2019.arxiv.org/abs/1910.04426
  26. [28]A.-K. Kassam and L. N. Trefethen. Fourth-order time-stepping for stiff pdes. SIAM Journal on Scientific Computing, 26(4):1214–1233, 2005.DOI
  27. [29]N. Kovachki, Z. Li, B. Liu, K. Azizzadeneshesheli, K. Bhattacharya, A. Stuart, and A. Anandakumar. Neural operator: Learning maps between function spaces with applications to PDEs. Journal of Machine Learning Research, 24(89):1–97, 2023.
  28. [30]A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, and M. W. Mahoney. Characterizing possible failure modes in physics-informed neural networks. Advances in Neural Information Processing Systems, 34:26548–26560, 2021.arxiv.org/abs/2109.01050
  29. [36]L.-J. Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine learning, 8(3):293–321, 1992.DOI
  30. [37]I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7.
  31. [39]R. J. McCann. A convexity principle for interacting gases. Advances in Mathematics, 128(1):153–179, 1997.
  32. [40]G. Mena and J. Niles-Weed. Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem. Advances in Neural Information Processing Systems, 32, 2019.
  33. [41]V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529–533, 2015.
  34. [43]J. Pathak, S. Subramanian, P. Harrington, S. Raja, A. Chattopadhyay, M. Mardani, T. Kurth, D. Hall, Z. Li, K. Azizzadenesheli, et al. FourCastNet: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214, 2022.
  35. [44]A.-A. Pooladian and J. Niles-Weed. Entropic estimation of optimal transport maps. arXiv preprint arXiv:2109.12004, 2021.arxiv.org/abs/2109.12004
  36. [45]M. Raissi, P. Perdikaris, and G. 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
  37. [46]P. J. Roache. Code verification by the method of manufactured solutions. J. Fluids Eng., 124(1):4–10, 2002.DOI
  38. [47]B. I. Shraiman, A. Pumir, W. van Saarloos, P. C. Hohenberg, H. Chaté, and M. Holen. Spatiotemporal chaos in the one-dimensional complex Ginzburg-Landau equation. Physica D: Nonlinear Phenomena, 57(3-4):241–248, 1992.
  39. [48]A. Sojitra and O. San. Method of manufactured learning for solver-free training of neural operators. arXiv preprint arXiv:2511.12890, 2025.arxiv.org/abs/2511.12890
  40. [49]P. Soviany, R. T. Ionescu, P. Rota, and N. Sebe. Curriculum learning: A survey. International Journal of Computer Vision, 130(6):1526–1565, 2022.arxiv.org/abs/2101.10382
  41. [51]C. Villani. Optimal transport: old and new, volume 338. Springer, 2009.
  42. [52]C. Wang, S. Li, D. He, and L. Wang. Is L2 physics informed loss always suitable for training physics informed neural network? Advances in Neural Information Processing Systems, 35:8278–8290, 2022a.arxiv.org/abs/2206.02016
  43. [53]C. Wang, J. Berner, Z. Li, D. Zhou, J. Wang, J. Bae, and A. Anandakumar. Coarse graining with neural operators for simulating chaotic systems. arXiv preprint arXiv:2408.05177, 2024a.arxiv.org/abs/2408.05177
  44. [54]H. Wang, B. Li, and H. Zhao. Understanding gradual domain adaptation: Improved analysis, optimal path and beyond. In International Conference on Machine Learning, pages 22784–22801. PMLR, 2022b.arxiv.org/abs/2204.08200
  45. [55]S. Wang, Y. Teng, and P. Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing, 43(5):A3055–A3081, 2021.
  46. [56]S. Wang, X. Yu, and P. Perdikaris. When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022c.arxiv.org/abs/2007.14527
  47. [57]S. Wang, S. Sankaran, and P. Perdikaris. Respecting causality for training physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 421:116813, 2024b.DOI
  48. [58]S. Wang, A. bhartari, B. Li, and P. Perdikaris. Gradient alignment in physics-informed neural networks: A second-order optimization perspective. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors, Advances in Neural Information Processing Systems, volume 38, Main Conference, pages 168482–168532. Curran Associates, Inc., 2025. doi: 10.52202/085713-5615. URL https://proceedings.neurips.cc/paper_files/paper/2025/file/f655706547885b8e32ef46f1c067ece2-Paper-Conference.pdf.DOI

Paper details

vlnc:W_q5jbsdnz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 13

downloads · 0

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