arxiv: 2605.09288 · v1 · submitted 2026-05-10 · 💻 cs.LG · cs.AI· cs.CE· cs.CV· cs.NA· math.NA

Recognition: no theorem link

MC²: Monte Carlo Correction for Fast Elliptic PDE Solving

Ethan Hsu , Hong Meng Yam , Ivan Ge

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:01 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CEcs.CVcs.NAmath.NA

keywords Monte CarloPDE solversneural correctionelliptic PDEsWalk-on-Sphereshybrid methodsbenchmark

0 comments

The pith

MC² corrects low-budget Monte Carlo PDE estimates with a neural network to match 1000× more compute accuracy

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

The paper establishes that errors in low-budget Monte Carlo solutions for elliptic PDEs are structured rather than purely random noise. A single neural network can learn to correct these errors and recover high-fidelity solutions. The resulting hybrid method combines the unbiased, geometry-agnostic properties of Walk-on-Spheres with learned speed. It achieves accuracy comparable to running over 1000 times more Monte Carlo samples while outperforming classical, denoising, and neural-operator baselines. The authors also release PDEZoo, a benchmark of 2 million elliptic PDEs with analytic ground truth, to support reproducible finite-compute studies.

Core claim

MC² treats a low-budget Monte Carlo solution as a structured estimator of the true field and learns a single-pass neural correction to recover a high-fidelity solution. This matches the accuracy of solutions using over 1000× more Monte Carlo compute and outperforms all evaluated classical, denoising, and neural-operator baselines across five elliptic PDE families and arbitrary geometric compositions.

What carries the argument

Hybrid WoS-NN solver that applies a learned neural correction to low-budget Walk-on-Spheres estimates of the solution field

Load-bearing premise

The structured error present in low-budget Monte Carlo solutions is learnable by a neural network that generalizes across five elliptic PDE families and arbitrary geometric compositions.

What would settle it

Failure of the neural correction to reach high-budget Monte Carlo accuracy on an elliptic PDE family or geometry outside the five families used for training would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.09288 by Ethan Hsu, Hong Meng Yam, Ivan Ge.

**Figure 1.** Figure 1: MC2 pipeline. A low-budget Monte Carlo WoS estimate is corrected in a single forward pass by a learned operator, yielding an improved solution for the PDE. in NLP and computer vision, where benchmarks like GLUE [35], SuperGLUE [36], and ImageNet [8] created a culture of head-to-head comparison on identical inputs. PDE solving has no analog. Existing benchmarks each occupy narrow regimes: PDEBench [34] and … view at source ↗

**Figure 2.** Figure 2: Overview of Data Distribution PDEZoo 4.1 Problem Specification Each PDE instance is defined by a tuple P = (Ω, E, f, g, u), where Ω ⊂ R d is a bounded domain with boundary ∂Ω, E denotes an elliptic PDE family, f : Ω → R is a forcing term, g : ∂Ω → R specifies Dirichlet boundary conditions, and u : Ω → R is the analytical solution. PDEZoo includes five elliptic PDE families: Laplace: ∆u = 0, Poisson: ∆u = f… view at source ↗

**Figure 3.** Figure 3: Qualitative comparison of MC2 against baselines on a Poisson instance. MC2 closely matches analytic ground truth and neural operators perform poorly. DiffusionPDE, CoCoGen, and raw WoS at B = 8 all retain substantial noise, and the best traditional denoiser over-smooths boundaries. evaluated using intermediate solutions. Budget sets, storage format, and reproducible generation details are in Appendix A.4. … view at source ↗

**Figure 4.** Figure 4: Scaling Study of MC2 (a) MC2 performance scales smoothly with training-set size at fixed input budget B = 8, with no sign of saturation. (b) We evaluate performance of applying MC2 error corrector to different WoS budgets. At every WoS budget, MC2 achieves substantially lower MSE than input WoS budget, matching the accuracy of ∼1,000× more MC compute. (c) On the accuracy–compute Pareto frontier, MC2 domina… view at source ↗

read the original abstract

Partial differential equation (PDE) solvers underpin scientific computing, but real-world deployment is bounded by compute. Classical Monte Carlo solvers such as Walk-on-Spheres (WoS) are unbiased and geometry-agnostic but are slow. Learned solvers are fast but biased and brittle under distribution shift. We present \textbf{MC$^2$}, a hybrid WoS-Neural Network (WoS-NN) PDE solver that treats a low-budget Monte Carlo solution as a structured estimator of the true field and learns a single-pass neural correction to recover a high-fidelity solution. MC$^2$ matches the accuracy of solutions using over $1000\times$ more Monte Carlo compute, outperforming all evaluated classical, denoising, and neural-operator baselines. To enable reproducible study of finite-compute PDE solving, we additionally release \textbf{PDEZoo}, the largest standardized elliptic PDE benchmark to date: 2M PDEs spanning five elliptic families and unlimited geometric compositions, with analytic ground truth and multi-budget Monte Carlo trajectories. Together \textbf{MC$^2$} and \textbf{PDEZoo} (1) empirically establish that finite-sample Monte Carlo error is structured, learnable, and correctable in a single forward pass, (2) show that we can solve PDEs $\sim$\textbf{1000x} faster than with just WoS, and (3) provide the evaluation infrastructure the field has so far lacked.

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

MC² pairs cheap Walk-on-Spheres with a neural correction and ships a large PDE benchmark, but the 1000x speedup rests on generalization that the stress test correctly flags as under-supported.

read the letter

The main takeaway is that low-budget Monte Carlo error for elliptic PDEs appears structured enough for a single network to correct it on the authors' test cases, delivering accuracy that would otherwise require far more samples. They back this with PDEZoo, a 2M-instance benchmark spanning five elliptic families and complex geometries, all with analytic ground truth and multi-budget trajectories. That dataset alone is a concrete step forward for anyone who has had to cobble together small test problems before.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces MC², a hybrid PDE solver that applies a neural network correction to low-budget Walk-on-Spheres Monte Carlo solutions for elliptic problems, claiming to recover accuracy equivalent to over 1000× more Monte Carlo samples while outperforming classical, denoising, and neural-operator baselines. It additionally releases PDEZoo, a benchmark of 2M elliptic PDE instances spanning five families with analytic ground truth and multi-budget trajectories, to support reproducible evaluation of finite-compute solvers.

Significance. If the central empirical claims hold, the work demonstrates that finite-sample Monte Carlo error in elliptic PDEs is sufficiently structured to be corrected in a single forward pass, offering a practical route to fast yet accurate solvers. The release of PDEZoo with standardized analytic ground truth and multi-budget trajectories is a clear strength that addresses the field's lack of reproducible benchmarks and could enable systematic study of error learnability across operators and geometries.

major comments (2)

[§5 (Experiments) and abstract] §5 (Experiments) and abstract: The headline claim that a single network trained on PDEZoo recovers high-fidelity solutions for arbitrary geometric compositions across five PDE families is load-bearing for the 1000× speedup and outperformance assertions, yet the reported tests do not include explicit out-of-distribution evaluation on topologies or coefficient heterogeneities materially different from the training distribution (e.g., multiply-connected domains or sharp coefficient jumps); without such controls the generalization risk identified in the stress-test note remains unaddressed.
[Table 3 (baseline comparison)] Table 3 (or equivalent baseline comparison table): The reported accuracy gains versus neural-operator and high-budget WoS baselines are presented without per-instance error distributions or statistical significance tests across the full 2M set; this makes it difficult to verify that the correction consistently closes the gap rather than succeeding only on average or on in-distribution subsets.

minor comments (2)

[§3 (Method)] §3 (Method): The precise encoding of the low-budget WoS field into the correction network (grid sampling, point cloud, or implicit representation) and any associated invariance properties should be stated explicitly to clarify how geometry-agnostic behavior is achieved.
[Figures] Figure captions: Several figures showing PDE solutions would benefit from explicit labels indicating the Monte Carlo sample budget used for the input field and the corresponding error norm relative to analytic ground truth.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and outline planned revisions to strengthen the empirical support for our claims.

read point-by-point responses

Referee: [§5 (Experiments) and abstract] The headline claim that a single network trained on PDEZoo recovers high-fidelity solutions for arbitrary geometric compositions across five PDE families is load-bearing for the 1000× speedup and outperformance assertions, yet the reported tests do not include explicit out-of-distribution evaluation on topologies or coefficient heterogeneities materially different from the training distribution (e.g., multiply-connected domains or sharp coefficient jumps); without such controls the generalization risk identified in the stress-test note remains unaddressed.

Authors: We acknowledge the referee's point that explicit OOD testing on multiply-connected domains and sharp coefficient discontinuities would more rigorously substantiate the generalization claims. PDEZoo's training and test splits already incorporate substantial geometric diversity and coefficient variations within the five families, and the stress-test note in the manuscript flags related risks. To directly address this, we will add a new subsection in §5 with dedicated OOD experiments on multiply-connected domains and sharp jumps, reporting errors relative to high-budget WoS and neural-operator baselines. revision: yes
Referee: [Table 3 (baseline comparison)] The reported accuracy gains versus neural-operator and high-budget WoS baselines are presented without per-instance error distributions or statistical significance tests across the full 2M set; this makes it difficult to verify that the correction consistently closes the gap rather than succeeding only on average or on in-distribution subsets.

Authors: We agree that per-instance distributions and formal significance testing would improve verifiability of consistent gains. The current Table 3 reports mean errors with standard deviations over the test sets. In revision we will expand the table and associated text to include per-instance error histograms (or violin plots) across the full 2M instances and add paired statistical tests (e.g., Wilcoxon signed-rank) to quantify that MC² improvements are significant and not confined to averages or in-distribution subsets. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical hybrid solver with independent benchmark validation

full rationale

The paper's derivation chain consists of an empirical proposal: train a neural network on low-budget Walk-on-Spheres trajectories paired with analytic ground truth from the released PDEZoo dataset to learn a single-pass correction. All headline claims (1000× speedup, outperformance vs. baselines) are established by direct comparison on held-out PDE instances across five families and arbitrary geometries, not by any equation that reduces to its own inputs, a fitted parameter renamed as prediction, or a self-citation chain. No uniqueness theorems, ansatzes, or renamings of known results are invoked; the work is self-contained against external benchmarks and does not rely on prior author results for its central premise.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical observation that Monte Carlo error is structured and correctable; this depends on standard well-posedness of elliptic PDEs and on the assumption that a neural network trained on the released benchmark will generalize.

free parameters (2)

Monte Carlo sample budget
The low budget used at inference time is a design choice that affects the input to the correction network.
neural network weights
The correction network is trained on PDEZoo data, introducing fitted parameters whose values are not reported in the abstract.

axioms (1)

standard math Elliptic PDEs admit unique solutions given suitable boundary conditions
Required for the Monte Carlo estimator and the notion of a true field to be well-defined.

pith-pipeline@v0.9.0 · 5572 in / 1251 out tokens · 41605 ms · 2026-05-12T04:01:08.306015+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

[1]

A unified deep artificial neural network approach to partial differential equations in complex geometries.Neurocomputing, 317:28–41, 2018

Jens Berg and Kaj Nyström. A unified deep artificial neural network approach to partial differential equations in complex geometries.Neurocomputing, 317:28–41, 2018

work page 2018
[2]

The rate of convergence of the Walk on Spheres algorithm

Ilia Binder and Mark Braverman. The rate of convergence of the Walk on Spheres algorithm. Geometric and Functional Analysis, 22(3):558–587, 2012. doi: 10.1007/s00039-012-0161-z

work page doi:10.1007/s00039-012-0161-z 2012
[3]

CRC Press, 2010

Mario Botsch, Leif Kobbelt, Mark Pauly, Pierre Alliez, and Bruno Lévy.Polygon Mesh Processing. CRC Press, 2010

work page 2010
[4]

Springer, 3rd edition, 2008

Susanne Brenner and Ridgway Scott.The Mathematical Theory of Finite Element Methods. Springer, 3rd edition, 2008

work page 2008
[5]

Briggs, Van Emden Henson, and Steve F

William L. Briggs, Van Emden Henson, and Steve F. McCormick.A Multigrid Tutorial. SIAM, 2nd edition, 2000

work page 2000
[6]

Digital geometry processing with discrete exterior calculus.ACM SIGGRAPH 2013 Courses, 2013

Keenan Crane, Fernando de Goes, Mathieu Desbrun, and Peter Schröder. Digital geometry processing with discrete exterior calculus.ACM SIGGRAPH 2013 Courses, 2013

work page 2013
[7]

Scientific machine learning through physics–informed neural networks: Where we are and what’s next.Journal of Scientific Computing, 92(3):88, 2022

Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next.Journal of Scientific Computing, 92(3):88, 2022

work page 2022
[8]

ImageNet:

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large- scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. doi: 10.1109/CVPR.2009.5206848

work page doi:10.1109/cvpr.2009.5206848 2009
[9]

Data science at the singularity.Harvard Data Science Review, 6(1), 2024

David Donoho. Data science at the singularity.Harvard Data Science Review, 6(1), 2024. doi: 10.1162/99608f92.b91339ef

work page doi:10.1162/99608f92.b91339ef 2024
[10]

SUNet: Swin transformer UNet for image denoising

Chi-Mao Fan, Tsung-Jung Liu, and Kuan-Hsien Liu. SUNet: Swin transformer UNet for image denoising. In2022 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, May 2022. doi: 10.1109/iscas48785.2022.9937486

work page doi:10.1109/iscas48785.2022.9937486 2022
[11]

Neural networks and the bias/variance dilemma.Neural Computation, 4(1):1–58, 1992

Stuart Geman, Elíe Bienenstock, and René Doursat. Neural networks and the bias/variance dilemma.Neural Computation, 4(1):1–58, 1992

work page 1992
[12]

Towards multi-spatiotemporal-scale generalized pde modeling.arXiv preprint arXiv:2209.15616, 2022

Jayesh K. Gupta and Johannes Brandstetter. Towards multi-spatiotemporal-scale generalized pde modeling.arXiv preprint arXiv:2209.15616, 2022

work page arXiv 2022
[13]

Springer, 2 edition, 2009

Trevor Hastie, Robert Tibshirani, and Jerome Friedman.The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer, 2 edition, 2009

work page 2009
[14]

Po- seidon: Efficient foundation models for pdes

Maximilian Herde, Nils Wandel, Siddhartha Mishra, Nils Thuerey, and Bogdan Raonic. Po- seidon: Efficient foundation models for pdes. InAdvances in Neural Information Processing Systems, 2024

work page 2024
[15]

Physics-embedded neural networks: Graph neural pde solvers with mixed boundary conditions

Masanobu Horie and Naoto Mitsume. Physics-embedded neural networks: Graph neural pde solvers with mixed boundary conditions. InAdvances in Neural Information Processing Systems, volume 35, pages 23218–23229. Curran Associates, Inc., 2022

work page 2022
[16]

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

work page 2024
[17]

Cocogen: Physically consistent and conditioned score-based generative models for forward and inverse problems.SIAM Journal on Scientific Computing, 2024

Christian Jacobsen, Yilin Zhuang, and Karthik Duraisamy. Cocogen: Physically consistent and conditioned score-based generative models for forward and inverse problems.SIAM Journal on Scientific Computing, 2024

work page 2024
[18]

A neural network-based pde solving algorithm with high precision.Scientific Reports, 13(1):4479, 2023

Zichao Jiang, Junyang Jiang, Qinghe Yao, and Gengchao Yang. A neural network-based pde solving algorithm with high precision.Scientific Reports, 13(1):4479, 2023

work page 2023
[19]

Monte carlo finite element method of structure reliability analysis.Reliability Engineering & System Safety, 40(1), 1993

Guoliang Jin, Lin Chen, and Jiamei Dong. Monte carlo finite element method of structure reliability analysis.Reliability Engineering & System Safety, 40(1), 1993. doi: 10.1016/ 0951-8320(93)90121-E. 10

work page 1993
[20]

Denoising diffusion restoration models

Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. InAdvances in Neural Information Processing Systems, volume 35, pages 23593–23606, 2022

work page 2022
[21]

Solving parametric pde problems with artificial neural networks.European Journal of Applied Mathematics, 32(3):421–435, 2021

Yuehaw Khoo, Jianfeng Lu, and Lexing Ying. Solving parametric pde problems with artificial neural networks.European Journal of Applied Mathematics, 32(3):421–435, 2021

work page 2021
[22]

Springer, 2 edition, 1998

Erich L Lehmann and George Casella.Theory of Point Estimation. Springer, 2 edition, 1998

work page 1998
[23]

Neural caches for monte carlo partial differential equation solvers

Zilu Li, Guandao Yang, Xi Deng, Christopher De Sa, Bharath Hariharan, and Steve Marschner. Neural caches for monte carlo partial differential equation solvers. InSIGGRAPH Asia 2023 Conference Papers, SA ’23. Association for Computing Machinery, 2023. doi: 10.1145/ 3610548.3618141

work page arXiv 2023
[24]

Fourier neural operator for parametric partial differential equations

Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. InInternational Conference on Learning Representations, 2021. URLhttps://openreview.net/forum?id=c8P9NQVtmnO

work page 2021
[25]

arXiv preprint arXiv:2111.03794 , year =

Zongyi Li, Hongkai Zheng, Nikola B. Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations.CoRR, abs/2111.03794, 2021. URL https://arxiv. org/abs/2111.03794

work page arXiv 2021
[26]

Walk On Spheres

Michael Mascagni and Chi-Ok Hwang. ϵ-Shell error analysis for “Walk On Spheres” algorithms. Mathematics and Computers in Simulation, 63(2):93–104, 2003. doi: 10.1016/S0378-4754(03) 00038-7

work page doi:10.1016/s0378-4754(03 2003
[27]

M. Muller. Some continuous monte carlo methods for the dirichlet problem.Ann. Math. Statist., 27(3), 1956

work page 1956
[28]

Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations, 2017

Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations, 2017

work page 2017
[29]

Challenges in training PINNs: A loss landscape perspective, 2024

Pratik Rathore, Weimu Lei, Zachary Frangella, Lu Lu, and Madeleine Udell. Challenges in training PINNs: A loss landscape perspective, 2024

work page 2024
[30]

Monte carlo geometry processing: A grid-free approach to pde-based methods on volumetric domains.ACM Trans

Rohan Sawhney and Keenan Crane. Monte carlo geometry processing: A grid-free approach to pde-based methods on volumetric domains.ACM Trans. Graph., 39(4), 2020

work page 2020
[31]

Sawhney, D

Rohan Sawhney, Dario Seyb, Wojciech Jarosz, and Keenan Crane. Grid-free monte carlo for pdes with spatially varying coefficients.ACM Trans. Graph., 41(4), July 2022. ISSN 0730-0301. doi: 10.1145/3528223.3530134

work page doi:10.1145/3528223.3530134 2022
[32]

Walk on stars: A grid-free monte carlo method for pdes with neumann boundary conditions.ACM Trans

Rohan Sawhney, Dario Seyb, Wojciech Jarosz, and Keenan Crane. Walk on stars: A grid-free monte carlo method for pdes with neumann boundary conditions.ACM Trans. Graph., 2023

work page 2023
[33]

A practical walk-on-boundary method for boundary value problems.ACM Transactions on Graphics, 42(4):1–16, July 2023

Ryusuke Sugimoto, Terry Chen, Yiti Jiang, Christopher Batty, and Toshiya Hachisuka. A practical walk-on-boundary method for boundary value problems.ACM Transactions on Graphics, 42(4):1–16, July 2023. ISSN 1557-7368. doi: 10.1145/3592109

work page doi:10.1145/3592109 2023
[34]

Pdebench: An extensive benchmark for scientific machine learning

Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning. InNeurIPS Datasets and Benchmarks Track, 2022

work page 2022
[35]

doi: 10.18653/v1/W18-5446

Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353–355, Brussels, Belgium, 2018. Association for Computational Linguis...

work page doi:10.18653/v1/w18-5446 2018
[36]

Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. SuperGLUE: A stickier benchmark for general-purpose language understanding systems. InAdvances in Neural Information Processing Systems, volume 32, 2019. 11

work page 2019
[37]

When and why pinns fail to train: A neural tangent kernel perspective.Journal of Computational Physics, 449:110768, 2022

Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective.Journal of Computational Physics, 449:110768, 2022

work page 2022
[38]

Monte carlo neural pde solver for learning pdes via probabilistic representation, 2023

Rui Zhang, Qi Meng, Rongchan Zhu, Yue Wang, Wenlei Shi, Shihua Zhang, Zhi-Ming Ma, and Tie-Yan Liu. Monte carlo neural pde solver for learning pdes via probabilistic representation, 2023. A PDEZoo: Detailed Specifications A.1 Solution Atom Types Each solution u(x, y) is constructed as a sum of n randomly chosen atoms, where n is sampled uniformly from [2,...

work page 2023