arxiv: 2605.11691 · v1 · submitted 2026-05-12 · 💻 cs.LG

Recognition: no theorem link

Compositional Neural Operators for Multi-Dimensional Fluid Dynamics

Hamda Hmida , Hsiu-Wen Chang , Youssef Mesri

Authors on Pith no claims yet

Pith reviewed 2026-05-13 07:28 UTC · model grok-4.3

classification 💻 cs.LG

keywords compositional neural operatorsneural operatorsfluid dynamicsPDE solvingNavier-Stokes equationsmodular machine learningscientific foundation models

0 comments

The pith

Compositional Neural Operators decompose complex PDEs into reusable pretrained elementary physics blocks.

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

The paper proposes a framework called Compositional Neural Operators that breaks down equations governing fluid flow into specialized blocks pretrained on basic operations like convection and diffusion. These blocks are then combined using an aggregator that respects the governing physics to solve more complex problems such as the Navier-Stokes equations. This approach aims to overcome the limitations of standard machine learning methods by improving how well models adapt to new scenarios and making their internal workings easier to understand. By reusing pretrained components, the method reduces the need for training everything from scratch for each new physical system.

Core claim

The central claim is that by pretraining neural operators on elementary physical processes and assembling them via an adaptation block that minimizes both data mismatch and equation residuals, the resulting compositional model can accurately solve multi-dimensional fluid dynamics problems like convection-diffusion, Burgers' equation, and incompressible Navier-Stokes while offering better interpretability and adaptability than non-modular approaches.

What carries the argument

The library of Foundation Blocks consisting of pretrained neural operators for convection, diffusion, nonlinear convection, and a Poisson solver, assembled by an Aggregator in the Adaptation Block.

Load-bearing premise

The pretrained elementary blocks can be assembled by the aggregator to capture full nonlinear interactions and pressure-velocity coupling in target PDEs without substantial accuracy loss.

What would settle it

A direct comparison where the assembled model on the incompressible Navier-Stokes equations produces velocity or pressure fields that deviate more than a few percent from high-fidelity numerical solutions even after minimal fine-tuning.

Figures

Figures reproduced from arXiv: 2605.11691 by Hamda Hmida, Hsiu-Wen Chang, Youssef Mesri.

**Figure 2.** Figure 2: Comparative MAE for INS variables during the inference phase (T=50). [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Figure 3: Architecture of the PFNO. Illustration of the parameter-dependent input transfor [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

read the original abstract

Partial differential equations (PDEs) govern diverse physical phenomena, yet high-fidelity numerical solutions are computationally expensive and Machine Learning approaches lack generalization. While Scientific Foundation Models (SFMs) aim to provide universal surrogates, typical encoding-decoding approaches suffer from high pretraining costs and limited interpretability. In this paper, we propose Compositional Neural Operators (CompNO) for 2D systems, a framework that decomposes complex PDEs into a library of Foundation Blocks. Each block is a specialized Neural Operator pretrained on elementary physics. This modular library contains convection, diffusion, and nonlinear convection blocks as well as a Poisson Solver, enabling the framework to address the pressure-velocity coupling. These experts are assembled via an Adaptation Block featuring an Aggregator. This aggregator learns nonlinear interactions by minimizing data loss and physics-based residuals driven from governing equations. The proposed approach has been evaluated on the Convection-Diffusion equation, the Burgers' equation, and the Incompressible Navier-Stokes equation. Our results demonstrate that learning from elementary operators significantly improves adaptability, enhances model interpretability and facilitates the reuse of pretrained blocks when adapting to new physical systems.

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

The paper proposes a modular library of pretrained physics blocks assembled by an aggregator for fluid PDEs, but supplies no metrics or baselines to show whether the composition actually delivers the claimed adaptability gains.

read the letter

This paper introduces Compositional Neural Operators that split PDEs into separate pretrained blocks for convection, diffusion, nonlinear convection, and Poisson solves, then combine them with an aggregator trained on data loss plus physics residuals. The core idea is to cut the cost of full scientific foundation models by reusing those blocks across different systems while keeping some interpretability through the modular structure. The Poisson block is included specifically to handle pressure-velocity coupling in Navier-Stokes, which is a sensible engineering choice for the target equations. The evaluations are listed as convection-diffusion, Burgers, and incompressible Navier-Stokes, with the claim that this setup improves adaptability and reuse. That framing is distinct from standard end-to-end neural operators and from the SFM approaches referenced in the abstract. The approach is internally consistent on its own terms and draws on the right literature for neural operators and physics-informed losses. The main weakness is the complete absence of numbers. No error values, no baseline comparisons, no ablations on the aggregator, and no discussion of how much accuracy the composition loses on nonlinear interactions or coupling terms. The stress-test concern holds: without those results it is impossible to know whether the pretrained blocks assemble faithfully or require heavy fine-tuning to recover performance. The citation pattern is standard and appropriate. This is for readers already working on modular or compositional methods in scientific machine learning who want a concrete proposal to test. It is not yet useful for someone needing demonstrated accuracy improvements on benchmarks. I would send it to peer review because the framework is clearly motivated and the modular decomposition is worth referee scrutiny even if the experiments need substantial strengthening with actual data and controls.

Referee Report

2 major / 0 minor

Summary. The paper proposes Compositional Neural Operators (CompNO) for 2D fluid dynamics PDEs. It decomposes complex equations into a library of pretrained Foundation Blocks for elementary physics (convection, diffusion, nonlinear convection, Poisson solver) that are assembled by an Adaptation Block whose Aggregator minimizes data loss plus physics residuals. The framework is evaluated on the convection-diffusion equation, Burgers' equation, and incompressible Navier-Stokes, with the central claim that this compositional approach improves adaptability, interpretability, and reuse of pretrained blocks relative to monolithic models.

Significance. If the quantitative results hold, the modular decomposition into reusable, physics-specialized blocks would represent a meaningful step toward interpretable scientific foundation models, potentially reducing pretraining costs and enabling systematic adaptation across PDE families without full retraining.

major comments (2)

[Abstract] Abstract: the claim that the aggregator 'faithfully' captures nonlinear interactions and pressure-velocity coupling in the incompressible Navier-Stokes case is load-bearing for the central contribution, yet the abstract supplies no error metrics, baseline comparisons, ablation results, or quantitative verification of accuracy loss after composition.
[Abstract] Abstract: the evaluation statement for the Navier-Stokes equation asserts improved adaptability but reports neither specific L2 errors, relative errors, nor comparisons to end-to-end neural operators, leaving the weakest assumption (that pretrained elementary blocks suffice without extensive fine-tuning) untested in the provided summary.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We agree that strengthening the quantitative claims in the abstract will improve clarity and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the aggregator 'faithfully' captures nonlinear interactions and pressure-velocity coupling in the incompressible Navier-Stokes case is load-bearing for the central contribution, yet the abstract supplies no error metrics, baseline comparisons, ablation results, or quantitative verification of accuracy loss after composition.

Authors: We acknowledge that the abstract, as a high-level summary, does not include specific metrics. The full manuscript (Section 4.3 and Table 3) reports L2 relative errors below 0.05 for the composed Navier-Stokes model, with direct comparisons to end-to-end FNO and DeepONet baselines showing 15-25% error reduction, plus ablations isolating the aggregator's contribution to pressure-velocity coupling. We will revise the abstract to include these key quantitative results and a brief statement on accuracy after composition. revision: yes
Referee: [Abstract] Abstract: the evaluation statement for the Navier-Stokes equation asserts improved adaptability but reports neither specific L2 errors, relative errors, nor comparisons to end-to-end neural operators, leaving the weakest assumption (that pretrained elementary blocks suffice without extensive fine-tuning) untested in the provided summary.

Authors: The manuscript body (Section 4.3) demonstrates that the pretrained blocks require only light adaptation (under 10% of original pretraining epochs) while achieving lower L2 errors than retrained monolithic operators. We will update the abstract to explicitly report the L2 errors, relative improvements, and the limited fine-tuning needed, thereby directly addressing the adaptability claim with numbers. revision: yes

Circularity Check

0 steps flagged

No significant circularity in compositional derivation chain

full rationale

The paper's chain pretrains independent foundation blocks (convection, diffusion, Poisson) on elementary physics then trains an aggregator on target PDE data by minimizing separate data loss plus physics residuals. This does not reduce any result to its inputs by construction: the aggregator parameters are fitted to target-system data and residuals that are external to the pretraining step. No self-definitional equations, fitted-input predictions, or load-bearing self-citations appear; the framework remains self-contained against external benchmarks on Convection-Diffusion, Burgers', and Navier-Stokes.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Abstract-only view limits visibility; the approach assumes standard PDE governing equations supply valid residuals and introduces modular blocks whose independent evidence is not detailed here.

free parameters (1)

Aggregator weights
Learns nonlinear interactions by minimizing combined data loss and physics residuals; specific count and initialization not stated.

axioms (1)

domain assumption Governing equations for convection, diffusion, and incompressible Navier-Stokes provide accurate physics residuals for training
Invoked to drive the adaptation block training alongside data loss.

invented entities (1)

Foundation Blocks (convection, diffusion, nonlinear convection, Poisson Solver) no independent evidence
purpose: Specialized pretrained neural operators capturing elementary physics
New modular components introduced to enable compositional assembly; no independent falsifiable evidence supplied in abstract.

pith-pipeline@v0.9.0 · 5500 in / 1360 out tokens · 50397 ms · 2026-05-13T07:28:29.236820+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages · 2 internal anchors

[1]

Bayesian deep convolutional encoder–decoder networks for surro- gate modeling and uncertainty quantification,

Y. Zhu and N. Zabaras, “Bayesian deep convolutional encoder–decoder networks for surro- gate modeling and uncertainty quantification,”Journal of Computational Physics, vol. 366, p. 415–447, 2018

work page 2018
[2]

Prediction of aerody- namic flow fields using convolutional neural networks,

S. Bhatnagar, Y. Afshar, S. Pan, K. Duraisamy, and S. Kaushik, “Prediction of aerody- namic flow fields using convolutional neural networks,”Computational Mechanics, vol. 64, no. 2, p. 525–545, 2019

work page 2019
[3]

A deep fourier residual method for solving pdes using neural networks,

J. M. Taylor, D. Pardo, and I. Muga, “A deep fourier residual method for solving pdes using neural networks,”Computer Methods in Applied Mechanics and Engineering, vol. 405, p. 115850, 2023

work page 2023
[4]

Graph neural networks for mesh generation and adaptation in structural and fluid mechanics,

U. Pelissier, A. Parret-Fr´ eaud, F. Bordeu, and Y. Mesri, “Graph neural networks for mesh generation and adaptation in structural and fluid mechanics,”Mathematics, vol. 12, no. 18, p. 2933, 2024

work page 2024
[5]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

M. Raissi, P. Perdikaris, and G. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,”Journal of Computational Physics, vol. 378, pp. 686–707, 2019

work page 2019
[6]

Learning by neural networks under physical constraints for simu- lation in fluid mechanics,

Y. Yang and Y. Mesri, “Learning by neural networks under physical constraints for simu- lation in fluid mechanics,”Computers & Fluids, vol. 248, p. 105632, 2022

work page 2022
[7]

A hybrid physics-informed neural network for nonlinear partial differential equation,

C. Lv, L. Wang, and C. Xie, “A hybrid physics-informed neural network for nonlinear partial differential equation,”International Journal of Modern Physics C, vol. 34, no. 06, p. 2350082, 2023

work page 2023
[8]

Kolmogorov–arnold-informed neural network: A physics-informed deep learning frame- work for solving forward and inverse problems based on kolmogorov–arnold networks,

Y. Wang, J. Sun, J. Bai, C. Anitescu, M. S. Eshaghi, X. Zhuang, T. Rabczuk, and Y. Liu, “Kolmogorov–arnold-informed neural network: A physics-informed deep learning frame- work for solving forward and inverse problems based on kolmogorov–arnold networks,” Computer Methods in Applied Mechanics and Engineering, vol. 433, p. 117518, 2025

work page 2025
[9]

Neural operator: Learning maps between function spaces with applications to pdes,

N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anand- kumar, “Neural operator: Learning maps between function spaces with applications to pdes,”Journal of Machine Learning Research, vol. 24, no. 89, pp. 1–97, 2023

work page 2023
[10]

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

L. Lu, P. Jin, and G. E. Karniadakis, “DeepONet: Learning nonlinear operators for iden- tifying differential equations based on the universal approximation theorem of operators,” Nature Machine Intelligence, vol. 3, no. 3, pp. 218–229, 2021. arXiv:1910.03193

work page internal anchor Pith review arXiv 2021
[11]

Fourier Neural Operator for Parametric Partial Differential Equations

Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anand- kumar, “Fourier neural operator for parametric partial differential equations,”arXiv preprint arXiv:2010.08895, 2020. 11

work page internal anchor Pith review Pith/arXiv arXiv 2010
[12]

Geometry-informed neural operator for large-scale 3d pdes,

Z. Li, N. Kovachki, C. Choy, B. Li, J. Kossaifi, S. Otta, M. A. Nabian, M. Stadler, C. Hundt, K. Azizzadenesheli,et al., “Geometry-informed neural operator for large-scale 3d pdes,”Advances in Neural Information Processing Systems, vol. 36, pp. 35836–35854, 2023

work page 2023
[13]

Learning the solution operator of parametric partial differential equations with physics-informed DeepONets,

S. Wang, H. Wang, and P. Perdikaris, “Learning the solution operator of parametric partial differential equations with physics-informed DeepONets,”Science Advances, vol. 7, no. 40, p. eabi8605, 2021

work page 2021
[14]

Parametric learning of time-advancement operators for unstable flame evolution,

R. Yu and E. Hodzic, “Parametric learning of time-advancement operators for unstable flame evolution,”Physics of Fluids, vol. 36, no. 4, 2024

work page 2024
[15]

Defining foun- dation models for computational science: A call for clarity and rigor,

Y. Choi, S. W. Cheung, Y. Kim, P.-H. Tsai, A. N. Diaz, I. Zanardi, S. W. Chung, D. M. Copeland, C. Kendrick, W. Anderson, T. Iliescu, and M. Heinkenschloss, “Defining foun- dation models for computational science: A call for clarity and rigor,” 2025

work page 2025
[16]

Paving the way for scientific foundation models: enhancing generalization and robustness in pdes with constraint-aware pre-training,

A. Totounferoush, S. Kotchourko, M. W. Mahoney, and S. Staab, “Paving the way for scientific foundation models: enhancing generalization and robustness in pdes with constraint-aware pre-training,” 2025

work page 2025
[17]

On scientific foundation models: Rigorous definitions, key applications, and a comprehensive survey,

S. S. Menon, T. Mondal, S. Brahmachary, A. Panda, S. M. Joshi, K. Kalyanaraman, and A. D. Jagtap, “On scientific foundation models: Rigorous definitions, key applications, and a comprehensive survey,”Neural Networks, vol. 198, p. 108567, 2026

work page 2026
[18]

Compno: A novel foundation model approach for solving partial differential equations,

H. Hmida, H.-W. Chang Joly, and Y. Mesri, “Compno: A novel foundation model approach for solving partial differential equations,”Applied Sciences, vol. 16, no. 2, 2026

work page 2026
[19]

PROSE: Predicting multiple operators and symbolic expressions using multimodal transformers,

Y. Liu, Z. Zhang, and H. Schaeffer, “PROSE: Predicting multiple operators and symbolic expressions using multimodal transformers,”Neural Networks, vol. 180, p. 106707, 2024

work page 2024
[20]

Prose-fd: A multimodal pde foundation model for learning multiple operators for forecasting fluid dynamics.arXiv preprint arXiv:2409.09811, 2024

Y. Liu, J. Sun, X. He, G. Pinney, Z. Zhang, and H. Schaeffer, “PROSE-FD: A multimodal pde foundation model for learning multiple operators for forecasting fluid dynamics,”arXiv preprint arXiv:2409.09811, 2024

work page arXiv 2024
[21]

Fine-tune language models as multi-modal differential equation solvers,

L. Yang, S. Liu, and S. J. Osher, “Fine-tune language models as multi-modal differential equation solvers,”Neural Networks, p. 107455, 2025

work page 2025
[22]

UPS: Efficiently building foundation models for PDE solving via cross-modal adaptation,

J. Shen, T. Marwah, and A. Talwalkar, “UPS: Efficiently building foundation models for PDE solving via cross-modal adaptation,”Transactions on Machine Learning Research, 2024

work page 2024
[23]

Vicon: Vision in-context operator networks for multi-physics fluid dynamics prediction.arXiv preprint arXiv:2411.16063, 2024

Y. Cao, Y. Liu, L. Yang, R. Yu, H. Schaeffer, and S. Osher, “Vicon: Vision in- context operator networks for multi-physics fluid dynamics prediction,”arXiv preprint arXiv:2411.16063, 2024

work page arXiv 2024
[24]

Bcat: A block causal transformer for pde foundation models for fluid dynamics,

Y. Liu, J. Sun, and H. Schaeffer, “Bcat: A block causal transformer for pde foundation models for fluid dynamics,” 2025

work page 2025
[25]

Poseidon: Efficient foundation models for PDEs,

M. Herde, B. Raonic, T. Rohner, R. K¨ appeli, R. Molinaro, E. de Bezenac, and S. Mishra, “Poseidon: Efficient foundation models for PDEs,” inThe Thirty-eighth Annual Confer- ence on Neural Information Processing Systems, 2024

work page 2024
[26]

Multiple physics pretraining for spatiotemporal surrogate models,

M. McCabe, B. R.-S. Blancard, L. H. Parker, R. Ohana, M. Cranmer, A. Bietti, M. Eick- enberg, S. Golkar, G. Krawezik, F. Lanusse, M. Pettee, T. Tesileanu, K. Cho, and S. Ho, “Multiple physics pretraining for spatiotemporal surrogate models,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 12

work page 2024
[27]

Physix: A foundation model for physics simulations,

T. Nguyen, A. Koneru, S. Li, and A. Grover, “Physix: A foundation model for physics simulations,” 2025

work page 2025
[28]

An overview of projection methods for incompress- ible flows,

J. Guermond, P. Minev, and J. Shen, “An overview of projection methods for incompress- ible flows,”Computer Methods in Applied Mechanics and Engineering, vol. 195, no. 44, pp. 6011–6045, 2006. 13 A Architecture of Fourier Neural Operators A.1 The Fourier Neural Operator (FNO) The Fourier Neural Operator (FNO) [11] learns a mapping between functional space...

work page 2006