arxiv: 2604.11375 · v2 · submitted 2026-04-13 · 🧮 math.NA · cs.NA

Recognition: unknown

DiLO: Decoupling Generative Priors and Neural Operators via Diffusion Latent Optimization for Inverse Problems

Guang Lin, Haibo Liu

Pith reviewed 2026-05-10 15:54 UTC · model grok-4.3

classification 🧮 math.NA cs.NA

keywords diffusion modelsinverse problemsneural operatorsPDE constraintsgenerative priorslatent optimizationmanifold consistencysurrogate solvers

0 comments

The pith

DiLO converts stochastic diffusion sampling into deterministic latent optimization to keep neural operators on physical manifolds for inverse problems.

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

The paper establishes that diffusion models provide flexible generative priors for PDE inverse problems but break when neural operator surrogates are applied to intermediate noisy states. DiLO solves this by turning the sampling process into an optimization trajectory over the initial latent variable, so the surrogate is always evaluated only on fully denoised physical parameters. This satisfies the manifold consistency requirement, permits stable gradient flow from measurements back to the latent start, and yields more accurate reconstructions. Experiments on electrical impedance tomography, inverse scattering, and inverse Navier-Stokes problems show gains in accuracy, speed, and noise robustness, backed by convergence theory.

Core claim

DiLO transforms the stochastic sampling process into a deterministic latent trajectory, enabling stable backpropagation of measurement gradients to the initial latent state. By keeping the trajectory on the physical manifold, it ensures physically valid updates and improves reconstruction accuracy while providing theoretical guarantees for convergence.

What carries the argument

Diffusion Latent Optimization (DiLO), which replaces stochastic diffusion sampling with deterministic optimization over the initial latent variable to enforce evaluation of neural surrogates exclusively on fully denoised physical states.

If this is right

Reconstruction accuracy improves because updates remain on the physical manifold at every step.
Measurement gradients propagate stably back to the initial latent without out-of-distribution evaluations.
The approach applies across Electrical Impedance Tomography, Inverse Scattering, and Inverse Navier-Stokes without retraining paired datasets.
Convergence of the latent trajectory is guaranteed under the stated theoretical conditions.

Where Pith is reading between the lines

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

The same latent-optimization idea could be tested on other generative priors such as flow-based models for the same inverse problems.
Running DiLO on experimental sensor data instead of synthetic measurements would check robustness outside simulated settings.
Extending the manifold consistency idea to time-evolving or high-dimensional PDEs not included in the current experiments could reveal further limits.

Load-bearing premise

Neural operator surrogates produce unreliable results when evaluated on partially denoised, non-physical intermediate states during diffusion sampling.

What would settle it

Demonstrating that a standard diffusion sampler with neural operators applied at every step achieves comparable accuracy and convergence on the same inverse problems without latent optimization.

Figures

Figures reproduced from arXiv: 2604.11375 by Guang Lin, Haibo Liu.

**Figure 1.** Figure 1: The DiLO framework. (A) Offline Training: Pre-training the Latent Diffusion Model (LDM) to serve as a plug-and-play prior, alongside a Differentiable Surrogate (e.g., FNO) acting as a forward physics operator. (B) Online Inference: Solving the inverse problem via deterministic latent optimization, where precise gradients from the measurement loss are backpropagated to the initial noise zT . where zt = √ α¯… view at source ↗

**Figure 2.** Figure 2: Detailed iterative reconstruction of the conductivity distribution for EIT. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

**Figure 3.** Figure 3: Detailed iterative reconstruction of the scattering contrast. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

**Figure 4.** Figure 4: Detailed iterative reconstruction of the initial vorticity field for the Inverse N-S problem. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

**Figure 5.** Figure 5: Qualitative comparison of reconstruction results. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

**Figure 6.** Figure 6: Qualitative comparison of reconstruction robustness under 50% Gaussian noise. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

**Figure 7.** Figure 7: Reconstruction analysis for a GT characterized by prominent, large-scale internal struc [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

**Figure 8.** Figure 8: Reconstruction analysis for a GT featuring multiple closely spaced, distinct anatomical [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗

**Figure 9.** Figure 9: Reconstruction analysis for an abdominal GT containing small, localized high-contrast [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

**Figure 10.** Figure 10: Reconstruction analysis for a chest GT exhibiting large, low-conductivity lung fields [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

**Figure 11.** Figure 11: Reconstruction analysis for sparse cell structures with heterogeneous internal contrasts, [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗

**Figure 12.** Figure 12: Reconstruction analysis for a tightly packed cluster of biological-like cells, confirming [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗

**Figure 13.** Figure 13: Reconstruction analysis for an initial vorticity field generated via a Gaussian random [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗

**Figure 14.** Figure 14: Reconstruction analysis for a GRF-generated vorticity field characterized by a highly [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗

**Figure 15.** Figure 15: Reconstruction analysis for a chaotic, multi-scale vorticity distribution sampled from [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗

**Figure 16.** Figure 16: Reconstruction analysis for a large-scale, coherent vortex structure originating from a [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗

**Figure 17.** Figure 17: Robustness analysis under 50% observation noise. [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗

read the original abstract

Diffusion models have emerged as powerful generative priors for solving PDE-constrained inverse problems. Compared to end-to-end approaches relying on massive paired datasets, explicitly decoupling the prior distribution of physical parameters from the forward physical model, a paradigm often formalized as Plug-and-Play (PnP) priors, offers enhanced flexibility and generalization. To accelerate inference within such decoupled frameworks, fast neural operators are employed as surrogate solvers. However, directly integrating them into standard diffusion sampling introduces a critical bottleneck: evaluating neural surrogates on partially denoised, non-physical intermediate states forces them into out-of-distribution (OOD) regimes. To eliminate this, the physical surrogate must be evaluated exclusively on the fully denoised parameter, a principle we formalize as the Manifold Consistency Requirement. To satisfy this requirement, we present Diffusion Latent Optimization (DiLO), which transforms the stochastic sampling process into a deterministic latent trajectory, enabling stable backpropagation of measurement gradients to the initial latent state. By keeping the trajectory on the physical manifold, it ensures physically valid updates and improves reconstruction accuracy. We provide theoretical guarantees for the convergence of this optimization trajectory. Extensive experiments across Electrical Impedance Tomography, Inverse Scattering, and Inverse Navier-Stokes problems demonstrate DiLO's accuracy, efficiency, and robustness to noise.

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

DiLO turns diffusion sampling into a deterministic latent optimization to keep neural operators on physical states during inverse problem solving.

read the letter

The core move here is spotting that standard PnP diffusion runs into trouble when neural operators see partially denoised states that are not physical. The authors formalize this as the Manifold Consistency Requirement and introduce DiLO to convert the usual stochastic process into a fixed trajectory over the initial latent variable. This lets them back-propagate measurement gradients while evaluating the surrogate only at the final denoised parameter, which stays on the manifold. Experiments on electrical impedance tomography, inverse scattering, and inverse Navier-Stokes show gains in accuracy and noise robustness compared with plain PnP diffusion. The claim of convergence guarantees for the trajectory is stated clearly in the abstract. That is the actual new piece: a concrete optimization wrapper that respects the physical surrogate constraint without retraining the diffusion model or the operator. The approach builds directly on existing decoupled prior work rather than inventing a wholly new generative framework, which keeps the scope manageable. The main soft spot is that the abstract alone does not show the proof details or the precise error bounds, so it is not yet clear how loose the guarantees become once discretization and approximation errors enter. The experiments use standard test cases; it would help to see whether the improvement holds when the forward model is more expensive or the latent dimension grows. Overall the construction is internally consistent and the problem it targets is real. This paper is aimed at researchers who already combine diffusion priors with fast PDE surrogates for inverse problems in computational science. It is worth sending to peer review because the method is simple enough to implement and the reported gains are on concrete applications that matter in the field.

Referee Report

1 major / 2 minor

Summary. The paper proposes Diffusion Latent Optimization (DiLO) to address a bottleneck when combining diffusion-based generative priors with neural operators for PDE-constrained inverse problems. It formalizes the Manifold Consistency Requirement (neural operators evaluated only on fully denoised physical parameters) and introduces a deterministic latent trajectory that replaces stochastic diffusion sampling. This enables stable backpropagation of measurement gradients to the initial latent state while preserving physical validity, with claimed theoretical convergence guarantees. Experiments on Electrical Impedance Tomography, Inverse Scattering, and Inverse Navier-Stokes demonstrate gains in accuracy, efficiency, and noise robustness compared to standard approaches.

Significance. If the central construction and convergence analysis hold, the work offers a principled decoupling of pre-trained generative priors from fast physical surrogates, avoiding OOD evaluation issues that plague direct integration. This could improve flexibility and generalization over end-to-end learned solvers while retaining the benefits of diffusion priors for ill-posed inverse problems. The explicit manifold constraint and theoretical guarantees distinguish it from heuristic PnP variants; reproducible code or parameter-free derivations would further strengthen its utility.

major comments (1)

[§4] §4 (theoretical analysis): The convergence guarantee for the DiLO trajectory is stated to follow from keeping updates on the physical manifold, but the proof sketch does not explicitly bound the deviation introduced by the neural operator approximation or the discretization of the latent trajectory; a quantitative error term relating the surrogate accuracy to the final reconstruction error would be needed to support the claim that DiLO improves accuracy over standard sampling.

minor comments (2)

[Abstract] Abstract: The phrase 'transforms the stochastic sampling process into a deterministic latent trajectory' is repeated; a single concise definition early in the abstract would improve readability.
[Experiments] Experiments section: The noise-robustness plots would benefit from error bars over multiple random seeds and a direct comparison table against a vanilla PnP baseline using the same neural operator but without the latent optimization step.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their positive summary, recognition of the work's potential significance, and recommendation for minor revision. We address the single major comment point-by-point below.

read point-by-point responses

Referee: [§4] §4 (theoretical analysis): The convergence guarantee for the DiLO trajectory is stated to follow from keeping updates on the physical manifold, but the proof sketch does not explicitly bound the deviation introduced by the neural operator approximation or the discretization of the latent trajectory; a quantitative error term relating the surrogate accuracy to the final reconstruction error would be needed to support the claim that DiLO improves accuracy over standard sampling.

Authors: We appreciate the referee's careful reading and agree that strengthening the error analysis would improve the presentation. The current proof in §4 establishes convergence of the deterministic latent optimization to a stationary point on the physical manifold under the manifold consistency requirement, treating the forward model as exact. In the revised manuscript we will augment §4 with an explicit quantitative error bound. The total reconstruction error will be decomposed as the sum of (i) the optimization error along the latent trajectory (which vanishes with increasing steps under the stated assumptions) and (ii) the neural-operator approximation error, controlled by the surrogate's Lipschitz constant restricted to the manifold. This bound will be contrasted with the additional out-of-distribution error incurred by standard diffusion sampling, thereby providing theoretical support for the accuracy gains observed in the experiments. The added analysis remains fully consistent with the existing convergence result and does not alter any claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper defines the Manifold Consistency Requirement as the principle that neural surrogates must be evaluated only on fully denoised physical parameters, then constructs DiLO as a deterministic latent trajectory optimization to satisfy it while enabling gradient backpropagation. This is a standard design choice for a new method, not a reduction of the claimed result to its own inputs by construction. No equations or steps are shown to equate a 'prediction' with a fitted parameter, no load-bearing self-citations reduce the central claim, and no ansatz or uniqueness theorem is smuggled in. The derivation remains self-contained with separate theoretical convergence guarantees claimed for the trajectory.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The approach relies on the newly introduced requirement and method; no free parameters or standard mathematical axioms are specified in the abstract.

axioms (1)

ad hoc to paper Manifold Consistency Requirement that neural operators are only evaluated on fully denoised physical parameters
Introduced to prevent OOD evaluation of surrogates during diffusion sampling.

invented entities (1)

Diffusion Latent Optimization (DiLO) no independent evidence
purpose: Transforms stochastic diffusion sampling into deterministic latent trajectory optimization
Proposed method to enable gradient backpropagation while maintaining physical validity.

pith-pipeline@v0.9.0 · 5521 in / 1120 out tokens · 32755 ms · 2026-05-10T15:54:59.338005+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 10 canonical work pages · 1 internal anchor

[1]

Electrical impedance tomography.Inverse problems, 18(6):R99, 2002

Liliana Borcea. Electrical impedance tomography.Inverse problems, 18(6):R99, 2002

2002
[2]

CRC press, 2004

David S Holder.Electrical impedance tomography: methods, history and applications. CRC press, 2004

2004
[3]

Springer, 1998

David L Colton, Rainer Kress, and Rainer Kress.Inverse acoustic and electromagnetic scat- tering theory, volume 93. Springer, 1998

1998
[4]

Extended contrast source inversion.Inverse problems, 15(5):1325, 1999

Peter M van den Berg, AL Van Broekhoven, and Aria Abubakar. Extended contrast source inversion.Inverse problems, 15(5):1325, 1999

1999
[5]

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

2019
[6]

SIAM, 2002

Curtis R Vogel.Computational methods for inverse problems. SIAM, 2002

2002
[7]

Walter De Gruyter, 2010

Anatoly B Bakushinsky, Mihail Yu Kokurin, and Alexandra Smirnova.Iterative methods for ill-posed problems: An introduction, volume 54. Walter De Gruyter, 2010

2010
[8]

Solving inverse obstacle scattering problem with latent surface representations.Inverse Problems, 40(6):065013, 2024

Junqing Chen, Bangti Jin, and Haibo Liu. Solving inverse obstacle scattering problem with latent surface representations.Inverse Problems, 40(6):065013, 2024. 20

2024
[9]

Deep learning based reconstruction meth- ods for electrical impedance tomography.arXiv preprint arXiv:2508.06281, 2025

Alexander Denker, Fabio Margotti, Jianfeng Ning, Kim Knudsen, Derick Nganyu Tanyu, Bangti Jin, Andreas Hauptmann, and Peter Maass. Deep learning based reconstruction meth- ods for electrical impedance tomography.arXiv preprint arXiv:2508.06281, 2025

work page arXiv 2025
[10]

Generative prior-guided neural interface recon- struction for 3d electrical impedance tomography.Journal of Computational Physics, page 114841, 2026

Haibo Liu, Junqing Chen, and Guang Lin. Generative prior-guided neural interface recon- struction for 3d electrical impedance tomography.Journal of Computational Physics, page 114841, 2026

2026
[11]

Denoising diffusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors,Advances in Neural Information Processing Systems, volume 33, pages 6840–6851. Curran Associates, Inc., 2020

2020
[12]

Score-based generative modeling through stochastic differential equations

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

2021
[13]

Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon

Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models.arXiv preprint arXiv:2111.08005, 2021

work page arXiv 2021
[14]

Diffusion posterior sampling for general noisy inverse problems

Hyungjin Chung, Jeongsol Kim, Michael Thompson Mccann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. InThe Eleventh International Conference on Learning Representations, 2023

2023
[15]

arXiv preprint arXiv:2305.04391 , year=

Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models.arXiv preprint arXiv:2305.04391, 2023

work page arXiv 2023
[16]

Inversebench: Benchmarking plug-and-play diffusion priors for inverse problems in phys- ical sciences

Hongkai Zheng, Wenda Chu, Bingliang Zhang, Zihui Wu, Austin Wang, Berthy Feng, Caifeng Zou, Yu Sun, Nikola Borislavov Kovachki, Zachary E Ross, Katherine Bouman, and Yisong Yue. Inversebench: Benchmarking plug-and-play diffusion priors for inverse problems in phys- ical sciences. InThe Thirteenth International Conference on Learning Representations, 2025

2025
[17]

Unleashing the denoising capability of diffusion prior for solving inverse problems.Advances in Neural Information Processing Systems, 37:47636–47677, 2024

Jiawei Zhang, Jiaxin Zhuang, Cheng Jin, Gen Li, and Yuantao Gu. Unleashing the denoising capability of diffusion prior for solving inverse problems.Advances in Neural Information Processing Systems, 37:47636–47677, 2024

2024
[18]

Deep data consistency: a fast and robust diffusion model-based solver for inverse problems.Neural Networks, page 108334, 2025

Hanyu Chen, Zhixiu Hao, and Liying Xiao. Deep data consistency: a fast and robust diffusion model-based solver for inverse problems.Neural Networks, page 108334, 2025

2025
[19]

Ode-dps: Ode-based diffusion posterior sampling for linear inverse problems in partial differential equation.Journal of Scientific Computing, 102(3):69, 2025

Enze Jiang, Jishen Peng, Zheng Ma, and Xiong-Bin Yan. Ode-dps: Ode-based diffusion posterior sampling for linear inverse problems in partial differential equation.Journal of Scientific Computing, 102(3):69, 2025

2025
[20]

Diffusion mod- els as probabilistic neural operators for recovering unobserved states of dynamical systems

Katsiaryna Haitsiukevich, Onur Poyraz, Pekka Marttinen, and Alexander Ilin. Diffusion mod- els as probabilistic neural operators for recovering unobserved states of dynamical systems. In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1–6. IEEE, 2024

2024
[21]

Dgno: A novel physics-aware neural op- erator for solving forward and inverse pde problems based on deep, generative probabilistic modeling.Generative Probabilistic Modeling, 2025

Yaohua Zang and Phaedon-Stelios Koutsourelakis. Dgno: A novel physics-aware neural op- erator for solving forward and inverse pde problems based on deep, generative probabilistic modeling.Generative Probabilistic Modeling, 2025

2025
[22]

Cocogen: Physically consistent and conditioned score-based generative models for forward and inverse problems.SIAM Jour- nal on Scientific Computing, 47(2):C399–C425, 2025

Christian Jacobsen, Yilin Zhuang, and Karthik Duraisamy. Cocogen: Physically consistent and conditioned score-based generative models for forward and inverse problems.SIAM Jour- nal on Scientific Computing, 47(2):C399–C425, 2025

2025
[23]

Physics-constrained flow matching: Sampling generative models with hard constraints.arXiv preprint arXiv:2506.04171, 2025

Utkarsh Utkarsh, Pengfei Cai, Alan Edelman, Rafael Gomez-Bombarelli, and Christopher Vin- cent Rackauckas. Physics-constrained flow matching: Sampling generative models with hard constraints.arXiv preprint arXiv:2506.04171, 2025

work page arXiv 2025
[24]

Invertible fourier neural operators for tackling both forward and inverse problems.arXiv preprint arXiv:2402.11722, 2024

Da Long, Zhitong Xu, Qiwei Yuan, Yin Yang, and Shandian Zhe. Invertible fourier neural operators for tackling both forward and inverse problems.arXiv preprint arXiv:2402.11722, 2024. 21

work page arXiv 2024
[25]

Physics-informed deep inverse operator networks for solving pde inverse problems.arXiv preprint arXiv:2412.03161, 2024

Sung Woong Cho and Hwijae Son. Physics-informed deep inverse operator networks for solving pde inverse problems.arXiv preprint arXiv:2412.03161, 2024

work page arXiv 2024
[26]

Diffusionpde: Generative pde-solving under partial observation.Advances in Neural Information Processing Systems, 37:130291–130323, 2024

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

2024
[27]

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B

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, abs/2505.17004, 2025

work page arXiv 2025
[28]

Y ., Yao, J., Chiang, L., Berner, J., and Anandkumar, A

Thomas YL Lin, Jiachen Yao, Lufang Chiang, Julius Berner, and Anima Anandkumar. Decoupled diffusion sampling for inverse problems on function spaces.arXiv preprint arXiv:2601.23280, 2026

work page arXiv 2026
[29]

Siam, 1990

David L Colton, Richard E Ewing, William Rundell, et al.Inverse problems in partial differ- ential equations, volume 42. Siam, 1990

1990
[30]

The calder´ on problem–an introduction to inverse problems.Preliminary notes on the book in preparation, 30, 2019

Joel Feldman, Mikko Salo, and Gunther Uhlmann. The calder´ on problem–an introduction to inverse problems.Preliminary notes on the book in preparation, 30, 2019

2019
[31]

Fourier Neural Operator for Parametric Partial Differential Equations

Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

work page internal anchor Pith review arXiv 2010
[32]

Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans.Medical physics, 38(2):9...

2011
[33]

Physics-informed neural operator for learning partial differential equations.ACM/IMS Journal of Data Science, 1(3):1–27, 2024

Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations.ACM/IMS Journal of Data Science, 1(3):1–27, 2024

2024
[34]

arXiv preprint arXiv:1910.03193 , year=

Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear opera- tors for identifying differential equations based on the universal approximation theorem of operators.arXiv preprint arXiv:1910.03193, 2019. 22 A Proof of Lemma 3.1 Proof.LetL exact(zT ) =L phys(M(zT )). The proof proceeds by evaluating the exact first- and second-order Fr...

work page arXiv 1910