pith. sign in

arxiv: 2607.01128 · v1 · pith:TVE7GWATnew · submitted 2026-07-01 · 💻 cs.LG · cs.NA· math.NA

GAIA: Geometry-Adaptive Operator Learning for Forward and Inverse Problems

Meenakshi Krishnan , Pranav Pulijala , Ke Chen , Haizhao Yang , Ramani Duraiswami This is my paper

Pith reviewed 2026-07-02 15:31 UTC · model grok-4.3

classification 💻 cs.LG cs.NAmath.NA
keywords operator learningPDE solversgeometry adaptiveintegral transformsforward problemsinverse problemsboundary value problemsneural operators
0
0 comments X

The pith

Encoding domain boundaries and fields into tokens lets one operator solve forward, BVP, and inverse PDE problems on arbitrary geometries in a single pass.

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

GAIA encodes the domain boundary and interior field distribution into geometry tokens. It conditions integral transform layers on these tokens through cross-attention so the kernel adapts locally to geometric features. This produces a single architecture that handles forward problems, including boundary value problems, and inverse problems on arbitrary domains without retraining, iterative optimization, or graph construction. The model is tested on seven 2D and 3D benchmarks, four of them new or extended for inverse and BVP settings such as electrical impedance tomography and a modified Poisson problem on mechanical components. It records new state-of-the-art accuracy on every inverse and BVP task while remaining competitive on standard forward problems and stable across point resolutions.

Core claim

By encoding the domain boundary and the interior field distribution into geometry tokens and conditioning integral transform layers on these tokens via cross-attention, GAIA yields a single architecture for forward (including BVPs) and inverse problems on arbitrary domains in one pass, without retraining, iterative optimization, or graph construction, and sets new state-of-the-art results on every inverse and BVP task.

What carries the argument

Geometry-Adaptive Integral Autoencoder (GAIA) that turns boundary and interior field information into tokens and uses cross-attention to condition integral transform layers for local kernel adaptation.

If this is right

  • GAIA reduces median relative L2 error by 64 percent on airfoil flow reconstruction and 27 percent on electrical impedance tomography relative to the next best amortized method.
  • The model outperforms all baselines on every shape category of the modified mechanical components benchmark for Poisson BVPs.
  • GAIA maintains stable accuracy across varying point resolutions where transformer-based baselines degrade.
  • The architecture solves both 2D and 3D problems without requiring graph construction or per-instance retraining.

Where Pith is reading between the lines

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

  • The token-based conditioning could allow rapid evaluation on families of geometries that differ only locally if the cross-attention mechanism generalizes beyond the training shapes.
  • Replacing the integral transform backbone with other kernel approximations might preserve the geometry adaptation while lowering memory cost for very large 3D domains.
  • The same encoding strategy might transfer to time-dependent or parametric PDEs if the tokens are extended to carry time or parameter information.

Load-bearing premise

Encoding the domain boundary and interior field distribution into geometry tokens and conditioning integral transform layers on these tokens via cross-attention is sufficient to adapt the kernel locally to geometric features for arbitrary domains and both forward and inverse problem types.

What would settle it

GAIA fails to reduce median relative L2 error below the next-best amortized method on a new inverse problem benchmark with an unseen arbitrary geometry, or its accuracy drops sharply on a BVP whose boundary conditions lie outside the token-encoded training distribution.

Figures

Figures reproduced from arXiv: 2607.01128 by Haizhao Yang, Ke Chen, Meenakshi Krishnan, Pranav Pulijala, Ramani Duraiswami.

Figure 1
Figure 1. Figure 1: GAIA architecture. Top: Two complementary tokenizers produce geometry tokens from boundary points and interior mesh-field values, refined by transformer encoder layers. Middle: A stack of geometry-conditioned Adaptive IAE blocks with DenseNet-style skip connections maps input features to the output field. Bottom: Each block follows an encode–process–decode pattern with integral-transform kernels conditione… view at source ↗
Figure 2
Figure 2. Figure 2: The two tokenizer pathways: GAIA encodes geometry through two complementary token sets. Top – Boundary tokenizer: Boundary point coordinates are processed by a PointNet-style encoder with shared MLPs and global max-pooling permutation-invariant tokens that summarize the global domain shape. Bottom – Slice tokenizer: The interior mesh coordinates and their associated field values are jointly embedded and as… view at source ↗
Figure 3
Figure 3. Figure 3: Model predictions on inverse benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results from the Mechanical Components Benchmark on the fitting, nut and gear categories. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Discretization invariance and noise robustness studies. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Inverse problems examples. (a) The EIT problem aims to reconstruct the internal con [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Model predictions on RTE. Comparison of ground truth solution against predictions from [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Model predictions on inverse benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Boundary Values on MCB shapes [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Results on the screws&bolts category from the Mechanical Components Benchmark. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Input output fields where ρ represents the mass density, u the displacement field, and σ the Cauchy stress tensor. The strain field is related to the displacement field via standard kinematic relations, and the system is closed by a hyperelastic constitutive model describing the non-linear stress-strain relationship. For this benchmark from [35], the stationary solution is considered for a unit square hyp… view at source ↗
Figure 12
Figure 12. Figure 12: Model predictions on Elasticity. Comparison of ground truth solution against predictions [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Model predictions on Poisson-Gauss. Comparison of ground truth solution against [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Discretization invariance on 3D Darcy. Although Transolver achieves lower error at the [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Discretization invariance on Poisson-Gauss. Multi-resolution augmentation (blue) provides [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
read the original abstract

Operator learning for partial differential equations (PDEs) on arbitrary geometries builds fast neural surrogates for large-scale simulation. Although recent geometry-adaptive neural operators have made substantial progress, they are mainly designed for forward problems in which inputs and outputs share the same spatial domain. This limits their applicability for boundary value problems (BVPs) and inverse problems, where inputs and outputs may live on different domains. We introduce the Geometry-Adaptive Integral Autoencoder (GAIA), an operator learning model that encodes the domain boundary and the interior field distribution into geometry tokens, and conditions integral transform layers on these tokens via cross-attention, allowing the kernel to adapt locally to geometric features. This yields a single architecture for forward (including BVPs) and inverse problems on arbitrary domains in one pass, without retraining, iterative optimization, or graph construction. We evaluate GAIA on seven 2D and 3D benchmarks, four of which are new or substantially extended benchmarks for inverse problems and BVP: electrical impedance tomography, optical tomography, 3D Darcy flow on varying geometries, and a modified setting of Poisson BVP on mechanical components benchmark (MCB). GAIA sets new state-of-the-art results on every inverse and BVP task, reducing median relative $L^2$ error by 64% on airfoil flow reconstruction and 27% on EIT relative to the next best amortized method, and outperforming all baselines on every shape category of MCB. On other forward problems, GAIA is competitive with specialized solvers while maintaining stable accuracy across point resolutions on which transformer-based baselines degrade.

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.

Referee Report

2 major / 2 minor

Summary. The paper introduces GAIA, a Geometry-Adaptive Integral Autoencoder that encodes domain boundary and interior field distribution into geometry tokens and conditions integral transform layers on these tokens via cross-attention. This produces a single architecture claimed to handle forward problems (including BVPs) and inverse problems on arbitrary domains in one pass, without retraining, iterative optimization, or graph construction. The model is evaluated on seven 2D and 3D benchmarks (four new or extended for inverse/BVP tasks: EIT, optical tomography, 3D Darcy on varying geometries, modified Poisson BVP on MCB), reporting new SOTA results on every inverse and BVP task with median relative L² error reductions of 64% on airfoil flow reconstruction and 27% on EIT relative to the next best amortized method, while remaining competitive on other forward problems.

Significance. If the reported empirical gains hold under rigorous controls, the work would be significant for extending geometry-adaptive neural operators beyond same-domain forward problems to unified handling of BVPs and inverse maps on arbitrary geometries. The introduction of new benchmarks for EIT, optical tomography, and modified MCB is a constructive contribution that could facilitate future comparisons in the field.

major comments (2)
  1. [§4] §4 (Experiments): the manuscript reports SOTA claims and specific error reductions (64% on airfoil, 27% on EIT) but provides no details on data splits, baseline re-implementations, number of runs, error bars, or statistical significance tests; these omissions are load-bearing because the central claim rests entirely on the benchmark comparisons.
  2. [§3] Method (cross-attention conditioning, §3): no capacity argument or analysis is given showing why cross-attention on geometry tokens suffices for local kernel adaptation when input and output supports are disjoint (BVPs, inverse problems); the claim of handling arbitrary domains without retraining therefore rests solely on the specific benchmark geometries tested.
minor comments (2)
  1. [Abstract] Abstract and §4: the statement of 'stable accuracy across point resolutions' lacks quantitative tables or figures showing the tested resolutions and direct comparison to transformer baselines that degrade.
  2. Notation for geometry tokens and integral transform conditioning could be introduced with an explicit diagram or equation reference earlier to improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential significance of GAIA for unified operator learning on arbitrary domains. We address the two major comments point by point below, indicating where revisions will be made.

read point-by-point responses

  1. Referee: [§4] §4 (Experiments): the manuscript reports SOTA claims and specific error reductions (64% on airfoil, 27% on EIT) but provides no details on data splits, baseline re-implementations, number of runs, error bars, or statistical significance tests; these omissions are load-bearing because the central claim rests entirely on the benchmark comparisons.

    Authors: We agree that these experimental details are necessary to rigorously support the reported performance improvements. In the revised version we will expand §4 with a dedicated subsection (and corresponding appendix) that specifies: (i) the exact train/validation/test splits for each of the seven benchmarks, (ii) the re-implementation protocol and hyper-parameter choices for all baselines, (iii) the number of independent random seeds used, (iv) mean ± standard deviation error bars, and (v) results of paired statistical significance tests (e.g., Wilcoxon signed-rank) between GAIA and the next-best amortized method on each task. This addition will make the empirical claims fully reproducible and verifiable. revision: yes

  2. Referee: [§3] Method (cross-attention conditioning, §3): no capacity argument or analysis is given showing why cross-attention on geometry tokens suffices for local kernel adaptation when input and output supports are disjoint (BVPs, inverse problems); the claim of handling arbitrary domains without retraining therefore rests solely on the specific benchmark geometries tested.

    Authors: We acknowledge that the manuscript does not supply a formal capacity or expressivity argument for the cross-attention mechanism. The design encodes both boundary and interior geometry into tokens that are then used to modulate the integral kernels via cross-attention; this allows each query point in the output domain to attend to geometric features regardless of whether the input and output supports coincide. The four new or extended benchmarks (EIT, optical tomography, 3D Darcy on varying geometries, modified Poisson BVP on MCB) deliberately include disjoint-support settings and diverse shape categories, and GAIA maintains consistent accuracy across them. To strengthen the presentation we will add a short paragraph in §3 that explains, at the level of the attention operation, why the conditioning enables local kernel adaptation on disjoint domains. We view this as a clarification rather than a new theoretical proof; a rigorous capacity analysis remains an interesting direction for future work. revision: partial

Circularity Check

0 steps flagged

No circularity; claims rest on empirical benchmarks

full rationale

The paper introduces the GAIA architecture for geometry-adaptive operator learning and supports its central claims (single-pass handling of forward/BVP/inverse problems on arbitrary domains, SOTA results) exclusively through empirical evaluation on seven 2D/3D benchmarks. No derivation chain, first-principles prediction, or mathematical reduction is presented that could reduce to fitted parameters, self-definitions, or self-citations by construction. The model description (geometry tokens + cross-attention conditioning of integral transforms) is an architectural choice whose performance is validated externally on held-out test cases, not derived from quantities defined within the paper itself. This is the standard non-circular pattern for empirical ML architecture papers.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The claim depends on the effectiveness of newly introduced model components (geometry tokens and cross-attention conditioning) that are not derived from first principles but designed and validated empirically; several architectural choices function as free parameters.

free parameters (2)
  • geometry token dimension
    Architectural hyperparameter chosen to represent boundary and field information for cross-attention.
  • number of cross-attention layers
    Model depth hyperparameter that controls how geometry information conditions the integral transforms.
axioms (1)
  • domain assumption Cross-attention on geometry tokens can produce locally adaptive kernels sufficient for arbitrary geometries and differing input/output domains.
    This is the core mechanistic assumption invoked to justify the unified architecture.
invented entities (1)
  • geometry tokens no independent evidence
    purpose: To encode domain boundary and interior field for conditioning the operator kernels.
    New representation introduced by the model with no independent evidence outside the empirical results.

pith-pipeline@v0.9.1-grok · 5842 in / 1367 out tokens · 41114 ms · 2026-07-02T15:31:13.264401+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 19 canonical work pages · 5 internal anchors

  1. [1]

    American Mathematical Society, 2024

    Roger Temam.Navier–Stokes equations: theory and numerical analysis, volume 343. American Mathematical Society, 2024

    2024

  2. [2]

    Electrical impedance tomography.Inverse problems, 18(6):R99–R136, 2002

    Liliana Borcea. Electrical impedance tomography.Inverse problems, 18(6):R99–R136, 2002

    2002

  3. [3]

    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

  4. [4]

    Regularization by architecture: A deep prior approach for inverse problems.Journal of Mathematical Imaging and Vision, 62(3):456–470, 2020

    Sören Dittmer, Tobias Kluth, Peter Maass, and Daniel Otero Baguer. Regularization by architecture: A deep prior approach for inverse problems.Journal of Mathematical Imaging and Vision, 62(3):456–470, 2020

    2020

  5. [5]

    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 differen- tial equations.arXiv preprint arXiv:2010.08895, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  6. [6]

    DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, and George Em Karniadakis. DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  7. [7]

    Geometry aware operator transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains.arXiv preprint arXiv:2505.18781, 2025

    Shizheng Wen, Arsh Kumbhat, Levi Lingsch, Sepehr Mousavi, Yizhou Zhao, Praveen Chan- drashekar, and Siddhartha Mishra. Geometry aware operator transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains.arXiv preprint arXiv:2505.18781, 2025

    work page arXiv 2025

  8. [8]

    Geometry-informed neural operator for large-scale 3D PDEs.Advances in Neural Information Processing Systems, 36:35836–35854, 2023

    Zongyi Li, Nikola Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Otta, Moham- mad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, et al. Geometry-informed neural operator for large-scale 3D PDEs.Advances in Neural Information Processing Systems, 36:35836–35854, 2023

    2023

  9. [9]

    Transolver: A Fast Transformer Solver for PDEs on General Geometries

    Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for PDEs on general geometries.arXiv preprint arXiv:2402.02366, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    Integral autoencoder network for discretization-invariant learning.Journal of Machine Learning Research, 23(286):1–45, 2022

    Yong Zheng Ong, Zuowei Shen, and Haizhao Yang. Integral autoencoder network for discretization-invariant learning.Journal of Machine Learning Research, 23(286):1–45, 2022

    2022

  11. [11]

    Operator learning with neural fields: Tackling PDEs on general geometries.Advances in Neural Information Processing Systems, 36:70581–70611, 2023

    Louis Serrano, Lise Le Boudec, Armand Kassaï Koupaï, Thomas X Wang, Yuan Yin, Jean-Noël Vittaut, and Patrick Gallinari. Operator learning with neural fields: Tackling PDEs on general geometries.Advances in Neural Information Processing Systems, 36:70581–70611, 2023

    2023

  12. [12]

    Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    2021

  13. [13]

    The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems.Communications in Mathematics and Statistics, 6(1):1–12, 2018

    Bing Yu et al. The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems.Communications in Mathematics and Statistics, 6(1):1–12, 2018

    2018

  14. [14]

    DGM: A deep learning algorithm for solving partial differential equations.Journal of computational physics, 375:1339–1364, 2018

    Justin Sirignano and Konstantinos Spiliopoulos. DGM: A deep learning algorithm for solving partial differential equations.Journal of computational physics, 375:1339–1364, 2018. 11

    2018

  15. [15]

    Gradient-enhanced physics- informed neural networks for forward and inverse PDE problems.Computer Methods in Applied Mechanics and Engineering, 393:114823, 2022

    Jeremy Yu, Lu Lu, Xuhui Meng, and George Em Karniadakis. Gradient-enhanced physics- informed neural networks for forward and inverse PDE problems.Computer Methods in Applied Mechanics and Engineering, 393:114823, 2022

    2022

  16. [16]

    Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM Journal on Scientific Computing, 43 (5):A3055–A3081, 2021

    Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM Journal on Scientific Computing, 43 (5):A3055–A3081, 2021

    2021

  17. [17]

    Bayesian deep convolutional encoder–decoder networks for surrogate modeling and uncertainty quantification.Journal of Computational Physics, 366: 415–447, 2018

    Yinhao Zhu and Nicholas Zabaras. Bayesian deep convolutional encoder–decoder networks for surrogate modeling and uncertainty quantification.Journal of Computational Physics, 366: 415–447, 2018

    2018

  18. [18]

    Prediction of aerodynamic flow fields using convolutional neural networks.Computational Mechanics, 64(2):525–545, 2019

    Saakaar Bhatnagar, Yaser Afshar, Shaowu Pan, Karthik Duraisamy, and Shailendra Kaushik. Prediction of aerodynamic flow fields using convolutional neural networks.Computational Mechanics, 64(2):525–545, 2019

    2019

  19. [19]

    Solving ill-posed inverse problems using iterative deep neural networks.Inverse Problems, 33(12):124007, 2017

    Jonas Adler and Ozan Öktem. Solving ill-posed inverse problems using iterative deep neural networks.Inverse Problems, 33(12):124007, 2017

    2017

  20. [20]

    Neural operator: Learning maps between function spaces with applications to PDEs.Journal of Machine Learning Research, 24(89):1–97, 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, 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

    2023

  21. [21]

    Factorized Fourier neural operators.arXiv preprint arXiv:2111.13802, 2021

    Alasdair Tran, Alexander Mathews, Lexing Xie, and Cheng Soon Ong. Factorized Fourier neural operators.arXiv preprint arXiv:2111.13802, 2021

    work page arXiv 2021

  22. [22]

    U-NO: U-shaped neural operators.arXiv preprint arXiv:2204.11127, 2022

    Md Ashiqur Rahman, Zachary E Ross, and Kamyar Azizzadenesheli. U-NO: U-shaped neural operators.arXiv preprint arXiv:2204.11127, 2022

    work page arXiv 2022

  23. [23]

    Multi-grid tensorized Fourier neural operator for high-resolution PDEs.arXiv preprint arXiv:2310.00120, 2023

    Jean Kossaifi, Nikola Kovachki, Kamyar Azizzadenesheli, and Anima Anandkumar. Multi-grid tensorized Fourier neural operator for high-resolution PDEs.arXiv preprint arXiv:2310.00120, 2023

    work page arXiv 2023

  24. [24]

    Neural Operator: Graph Kernel Network for Partial Differential Equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations.arXiv preprint arXiv:2003.03485, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2003

  25. [25]

    Multipole graph neural operator for parametric partial differential equations.Advances in Neural Information Processing Systems, 33:6755–6766, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Andrew Stuart, Kaushik Bhattacharya, and Anima Anandkumar. Multipole graph neural operator for parametric partial differential equations.Advances in Neural Information Processing Systems, 33:6755–6766, 2020

    2020

  26. [26]

    Wavelet neural operator: a neural operator for parametric partial differential equations.arXiv preprint arXiv:2205.02191, 2022

    Tapas Tripura and Souvik Chakraborty. Wavelet neural operator: a neural operator for parametric partial differential equations.arXiv preprint arXiv:2205.02191, 2022

    work page arXiv 2022

  27. [27]

    Learning to accelerate partial differential equations via latent global evolution.Advances in Neural Information Processing Systems, 35: 2240–2253, 2022

    Tailin Wu, Takashi Maruyama, and Jure Leskovec. Learning to accelerate partial differential equations via latent global evolution.Advances in Neural Information Processing Systems, 35: 2240–2253, 2022

    2022

  28. [28]

    Uncertainty quantification for forward and inverse problems of PDEs via latent global evolution

    Tailin Wu, Willie Neiswanger, Hongtao Zheng, Stefano Ermon, and Jure Leskovec. Uncertainty quantification for forward and inverse problems of PDEs via latent global evolution. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 320–328, 2024

    2024

  29. [29]

    Derivative-informed neural operator acceleration of geometric MCMC for infinite-dimensional Bayesian inverse problems

    Lianghao Cao, Thomas O’Leary-Roseberry, and Omar Ghattas. Derivative-informed neural operator acceleration of geometric MCMC for infinite-dimensional Bayesian inverse problems. Journal of Machine Learning Research, 26(78):1–68, 2025

    2025

  30. [30]

    Semi-supervised invertible neural operators for Bayesian inverse problems.Computational Mechanics, 72(3): 451–470, 2023

    Sebastian Kaltenbach, Paris Perdikaris, and Phaedon-Stelios Koutsourelakis. Semi-supervised invertible neural operators for Bayesian inverse problems.Computational Mechanics, 72(3): 451–470, 2023. 12

    2023

  31. [31]

    On conditional diffusion mod- els for PDE simulations.Advances in Neural Information Processing Systems, 37:23246–23300, 2024

    Aliaksandra Shysheya, Cristiana Diaconu, Federico Bergamin, Paris Perdikaris, José M Hernández-Lobato, Richard E Turner, and Emile Mathieu. On conditional diffusion mod- els for PDE simulations.Advances in Neural Information Processing Systems, 37:23246–23300, 2024

    2024

  32. [32]

    Neural inverse operators for solving PDE inverse problems.arXiv preprint arXiv:2301.11167, 2023

    Roberto Molinaro, Yunan Yang, Björn Engquist, and Siddhartha Mishra. Neural inverse operators for solving PDE inverse problems.arXiv preprint arXiv:2301.11167, 2023

    work page arXiv 2023

  33. [33]

    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

    work page arXiv 2024

  34. [34]

    Operator learning meets inverse problems: A probabilistic perspective.arXiv preprint arXiv:2508.20207, 2025

    Nicholas H Nelsen and Yunan Yang. Operator learning meets inverse problems: A probabilistic perspective.arXiv preprint arXiv:2508.20207, 2025

    work page arXiv 2025

  35. [35]

    Fourier neural operator with learned deformations for PDEs on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023

    Zongyi Li, Daniel Zhengyu Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for PDEs on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023

    2023

  36. [36]

    RIGNO: A graph-based framework for robust and accurate operator learning for PDEs on arbitrary domains.arXiv preprint arXiv:2501.19205, 2025

    Sepehr Mousavi, Shizheng Wen, Levi Lingsch, Maximilian Herde, Bogdan Raoni´c, and Sid- dhartha Mishra. RIGNO: A graph-based framework for robust and accurate operator learning for PDEs on arbitrary domains.arXiv preprint arXiv:2501.19205, 2025

    work page arXiv 2025

  37. [37]

    Neural fields for rapid aircraft aerodynamics simulations.Scientific Reports, 14 (1):25496, 2024

    Giovanni Catalani, Siddhant Agarwal, Xavier Bertrand, Frederic Tost, Michael Bauerheim, and Joseph Morlier. Neural fields for rapid aircraft aerodynamics simulations.Scientific Reports, 14 (1):25496, 2024

    2024

  38. [38]

    Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communi- cations, 15(1):10416, 2024

    Pan Du, Meet Hemant Parikh, Xiantao Fan, Xin-Yang Liu, and Jian-Xun Wang. Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communi- cations, 15(1):10416, 2024

    2024

  39. [39]

    Latent neural operator for solving forward and inverse PDE problems.Advances in Neural Information Processing Systems, 37:33085–33107, 2024

    Tian Wang and Chuang Wang. Latent neural operator for solving forward and inverse PDE problems.Advances in Neural Information Processing Systems, 37:33085–33107, 2024

    2024

  40. [40]

    GNOT: A general neural operator transformer for operator learning

    Zhongkai Hao, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song, and Jun Zhu. GNOT: A general neural operator transformer for operator learning. InInternational conference on machine learning, pages 12556–12569. PMLR, 2023

    2023

  41. [41]

    Transformer for partial differential equations’ operator learning.arXiv preprint arXiv:2205.13671, 2022

    Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differential equations’ operator learning.arXiv preprint arXiv:2205.13671, 2022

    work page arXiv 2022

  42. [42]

    Improved operator learning by orthogonal attention.arXiv preprint arXiv:2310.12487, 2023

    Zipeng Xiao, Zhongkai Hao, Bokai Lin, Zhijie Deng, and Hang Su. Improved operator learning by orthogonal attention.arXiv preprint arXiv:2310.12487, 2023

    work page arXiv 2023

  43. [43]

    Neural Green’s functions

    Seungwoo Yoo, Kyeongmin Yeo, Jisung Hwang, and Minhyuk Sung. Neural Green’s functions. arXiv preprint arXiv:2511.01924, 2025

    work page arXiv 2025

  44. [44]

    PointNet: Deep learning on point sets for 3D classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017

    2017

  45. [45]

    NeRF: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoor- thi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021

    2021

  46. [46]

    On an inverse boundary value problem.Computational & Applied Mathematics, 25:133–138, 2006

    Alberto P Calderón. On an inverse boundary value problem.Computational & Applied Mathematics, 25:133–138, 2006

    2006

  47. [47]

    Solving optical tomography with deep learning.arXiv preprint arXiv:1910.04756, 2019

    Yuwei Fan and Lexing Ying. Solving optical tomography with deep learning.arXiv preprint arXiv:1910.04756, 2019. 13

    work page arXiv 1910

  48. [48]

    A large-scale annotated mechanical components benchmark for classification and retrieval tasks with deep neural networks

    Sangpil Kim, Hyung-gun Chi, Xiao Hu, Qixing Huang, and Karthik Ramani. A large-scale annotated mechanical components benchmark for classification and retrieval tasks with deep neural networks. InEuropean conference on computer vision, pages 175–191. Springer, 2020

    2020

  49. [49]

    Springer Science & Business Media, 2012

    Anders Logg, Kent-Andre Mardal, and Garth Wells.Automated solution of differential equations by the finite element method: The FEniCS book, volume 84. Springer Science & Business Media, 2012

    2012

  50. [50]

    Transformers are rnns: Fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. InInternational conference on machine learning, pages 5156–5165. PMLR, 2020

    2020

  51. [51]

    EIT reconstruction algorithms: pitfalls, challenges and recent develop- ments.Physiological measurement, 25(1):125, 2004

    William RB Lionheart. EIT reconstruction algorithms: pitfalls, challenges and recent develop- ments.Physiological measurement, 25(1):125, 2004

    2004

  52. [52]

    Inverse transport theory of photoacous- tics.Inverse Problems, 26(2):025011, 2010

    Guillaume Bal, Alexandre Jollivet, and Vincent Jugnon. Inverse transport theory of photoacous- tics.Inverse Problems, 26(2):025011, 2010

    2010

  53. [53]

    Elsevier, 2014

    Gerald Farin.Curves and surfaces for computer-aided geometric design: a practical guide. Elsevier, 2014

    2014

  54. [54]

    Fast tetrahedral meshing in the wild

    Yixin Hu, Teseo Schneider, Bolun Wang, Denis Zorin, and Daniele Panozzo. Fast tetrahedral meshing in the wild. 2020

    2020

  55. [55]

    Adam: A Method for Stochastic Optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014. A Datasets and Benchmarks Below is a brief summary of the datasets that we use in the paper. A.1 Inverse Problem: Electrical Impedance Tomography Electrical Impedance Tomography (EIT) is an imaging technique that involves solving an inverse pro...

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  56. [56]

    3D Poisson BVP[ 43] −∆u=f analytical, u|∂Ω =g(x, y, z)

    4096 train / 2048 test samples. 3D Poisson BVP[ 43] −∆u=f analytical, u|∂Ω =g(x, y, z) . Os- cillatory analytical source; randomized poly- nomial Dirichlet BCs. BVP / Forward.3D mechanical parts (MCB): gears, nuts, fittings, screws & bolts. Setting adapted from [43]. 200 shapes/category train, 20 test with 50 boundary conditions per shape. EIT[32]∇ ·(a(x)...

    2048

  57. [57]

    The scattering field is defined as: σs(x) = 1.0 + 1 + exp −3 mX k=1 ck sin(ωx,kπx′

    denote the spatial coordinates normalized by the domain scale factor. The scattering field is defined as: σs(x) = 1.0 + 1 + exp −3 mX k=1 ck sin(ωx,kπx′

  58. [58]

    star shape

    sin(ωy,kπx′ 2) !!−1 , 16 where the number of modes m is drawn uniformly from {2, . . . ,5}, the expansion coefficients ck are sampled independently from a uniform distribution U([−1,1]) , and the spatial frequencies ωx,k and ωy,k are drawn uniformly from the integer set {1,2,3} . This construction ensures that the scattering coefficients are smooth and st...

    2048

  59. [59]

    GAIA achieves competitive latency and memory while maintaining the highest throughput among geometry-adaptive methods

    after 20 warm-up iterations with CUDA synchronization; throughput is measured at batch size 16. GAIA achieves competitive latency and memory while maintaining the highest throughput among geometry-adaptive methods. To assess the computational efficiency of GAIA relative to baseline methods, we conduct a compre- hensive timing analysis on the EIT benchmark...

    work page arXiv 2000