pith. machine review for the scientific record. sign in

arxiv: 2604.20993 · v1 · submitted 2026-04-22 · 💻 cs.LG

Recognition: unknown

Droplet-LNO: Physics-Informed Laplace Neural Operators for Accurate Prediction of Droplet Spreading Dynamics on Complex Surfaces

Ganesh Sahadeo Meshram , Partha Pratim Chakrabarti , Suman Chakraborty
Authors on Pith no claims yet

Pith reviewed 2026-05-10 01:02 UTC · model grok-4.3

classification 💻 cs.LG
keywords droplet spreadingphysics-informed neural operatorsLaplace transformsmultiphysics simulationCahn-Hilliard equationNavier-Stokes equationsneural operator learningsurface wetting
0
0 comments X

The pith

A physics-informed Laplace neural operator learns to predict droplet spreading dynamics on varied surfaces by embedding exponential transient behavior directly into its architecture.

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

The paper presents PI-LNO as a neural operator that treats the Laplace integral transform as a trainable functional basis to capture the time evolution of liquid droplets on solid substrates. Conventional CFD methods require 18 to 24 hours for each transient simulation of this multiphysics process, limiting their use in design loops for printing, cooling, and microfluidic devices. Training occurs on CFD-generated data covering contact angles from 20 to 160 degrees, with a composite loss that penalizes deviations from data while enforcing Navier-Stokes flow, Cahn-Hilliard phase separation, and causality rules. The resulting model aims to deliver fast, physically consistent forecasts that generalize across surface types without repeated full simulations.

Core claim

Through the Laplace integral transform serving as a learned physics-informed functional basis, PI-LNO natively represents the exponential transient dynamics of droplet spreading. When trained with a physics-regularized loss that combines data fidelity terms (MSE, MAE, RMSE) and constraints drawn from the Navier-Stokes and Cahn-Hilliard equations plus causality conditions, the architecture produces predictions that remain consistent with the underlying multiphysics on complex surfaces.

What carries the argument

The Laplace integral transform function, used as a learned physics-informed functional basis that directly encodes the exponential time-domain decay and transient response of the spreading process.

If this is right

  • Droplet morphology and contact-line motion can be obtained in seconds rather than hours for each new surface condition.
  • Physical consistency with fluid dynamics and phase-field equations holds across the full range of tested contact angles without separate correction steps.
  • Rapid exploration of spreading behavior becomes feasible for applications such as inkjet printing and spray cooling on varied substrates.
  • The same operator architecture can be retrained on additional CFD datasets to extend coverage to new surface textures or fluid properties.

Where Pith is reading between the lines

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

  • The same Laplace-basis construction could be applied to other transient multiphysics problems that exhibit exponential relaxation, such as certain heat-transfer or viscoelastic flows.
  • If inference remains stable, the model might support real-time feedback loops in microfluidic control systems where full CFD is too slow.
  • Operator learning that incorporates integral transforms may offer a route to respect causality and decay properties in a wider class of physical systems without explicit time-stepping.

Load-bearing premise

That embedding Navier-Stokes, Cahn-Hilliard, and causality constraints inside a single composite loss is sufficient to enforce physical consistency and generalization across surfaces without hidden data-specific biases or post-training corrections.

What would settle it

Generate a new CFD simulation for an unseen surface geometry with contact angle inside the training range, run the trained PI-LNO forward pass, and check whether the predicted velocity and pressure fields satisfy the Navier-Stokes equations to within discretization error; repeated large violations would falsify the claim of enforced physical consistency.

read the original abstract

Spreading of liquid droplets on solid substrates constitutes a classic multiphysics problem with widespread applications ranging from inkjet printing, spray cooling, to biomedical microfluidic systems. Yet, accurate computational fluid dynamic (CFD) simulations are prohibitively expensive, taking more than 18 to 24 hours for each transient computation. In this paper, Physics-Informed Laplace Operator Neural Network (PI-LNO) is introduced, representing a novel architecture where the Laplace integral transform function serves as a learned physics-informed functional basis. Extensive comparative benchmark studies were performed against five other state-of-the-art approaches: UNet, UNet with attention modules (UNet-AM), DeepONet, Physics-Informed UNet (PI-UNet), and Laplace Neural Operator (LNO). Through complex Laplace transforms, PI-LNO natively models the exponential transient dynamics of the spreading process. A TensorFlow-based PI-LNO is trained on multi-surface CFD data spanning contact angles $\theta_s \epsilon [20,160]$, employing a physics-regularized composite loss combining data fidelity (MSE, MAE, RMSE) with Navier-Stokes, Cahn-Hilliard, and causality constraints.

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

3 major / 1 minor

Summary. The manuscript introduces the Physics-Informed Laplace Neural Operator (PI-LNO), a neural operator architecture that employs the Laplace integral transform as a learned physics-informed functional basis for predicting the transient spreading dynamics of liquid droplets on solid substrates with varying contact angles. The model is trained on multi-surface CFD data for contact angles in [20,160] using a composite loss that combines data-fidelity terms (MSE, MAE, RMSE) with constraints from the Navier-Stokes and Cahn-Hilliard equations plus causality. It is benchmarked against UNet, UNet-AM, DeepONet, PI-UNet, and LNO, with the central claim that complex Laplace transforms allow PI-LNO to natively capture the exponential transient behavior of the spreading process.

Significance. If the accuracy and generalization claims are substantiated with quantitative evidence, the work could meaningfully accelerate multiphysics simulations of droplet dynamics that currently require 18-24 hours per CFD run, with relevance to inkjet printing, spray cooling, and microfluidics. The integration of a Laplace-based functional basis with physics constraints represents a potentially useful direction for operator learning on transient nonlinear PDE systems, though the manuscript does not yet demonstrate that this basis independently encodes the target dynamics.

major comments (3)
  1. [Abstract and §1] Abstract and §1 (central claim): The assertion that 'Through complex Laplace transforms, PI-LNO natively models the exponential transient dynamics of the spreading process' is not supported by any derivation showing how the Laplace integral transform produces exponential solutions for the nonlinear coupled Navier-Stokes + Cahn-Hilliard system. The exponential character could arise from the data term or the composite loss rather than the transform layer itself; a concrete mapping from the learned Laplace basis to the transient form of the droplet equations is required.
  2. [Abstract and benchmark section] Abstract and benchmark section: The manuscript states that 'extensive comparative benchmark studies' were performed against five baselines yet supplies no quantitative error metrics (e.g., L2 errors, relative errors, or convergence rates), no tables of performance numbers, and no mention of training/validation splits or generalization across surfaces. Without these data it is impossible to verify that PI-LNO supports the accuracy claims or outperforms the baselines in a load-bearing way.
  3. [Training and loss description] Training and loss description: The physics-regularized composite loss is described as combining data fidelity with Navier-Stokes, Cahn-Hilliard, and causality constraints, but no explicit weighting schedule, enforcement mechanism, or ablation study is provided to demonstrate that the physics terms enforce consistency independently of the CFD training data. This leaves open the possibility that outputs reduce to data-driven fits rather than independent physics-informed predictions.
minor comments (1)
  1. [Abstract] Notation: The contact-angle range is written as '$θ_s ε [20,160]$'; this should be corrected to the standard set notation '$θ_s ∈ [20,160]$'.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough and constructive review of our manuscript on the Physics-Informed Laplace Neural Operator (PI-LNO) for droplet spreading dynamics. We address each major comment point by point below, providing clarifications and committing to revisions that strengthen the substantiation of our claims without misrepresenting the work.

read point-by-point responses

  1. Referee: [Abstract and §1] Abstract and §1 (central claim): The assertion that 'Through complex Laplace transforms, PI-LNO natively models the exponential transient dynamics of the spreading process' is not supported by any derivation showing how the Laplace integral transform produces exponential solutions for the nonlinear coupled Navier-Stokes + Cahn-Hilliard system. The exponential character could arise from the data term or the composite loss rather than the transform layer itself; a concrete mapping from the learned Laplace basis to the transient form of the droplet equations is required.

    Authors: We acknowledge that the manuscript would benefit from a more explicit explanation of how the complex Laplace transform serves as a functional basis for the exponential transients in this nonlinear system. The transform is chosen because it maps decaying exponential behaviors to algebraic forms in the s-domain, enabling efficient representation of the spreading dynamics; for the coupled nonlinear PDEs this is realized through the learned operator combined with the physics loss. To address the concern directly, we will revise the abstract slightly for precision and add a new paragraph in the methods section with a step-by-step rationale, including the linearised droplet equation case where the mapping is exact, while noting the role of data and loss in the full nonlinear setting. revision: yes

  2. Referee: [Abstract and benchmark section] Abstract and benchmark section: The manuscript states that 'extensive comparative benchmark studies' were performed against five baselines yet supplies no quantitative error metrics (e.g., L2 errors, relative errors, or convergence rates), no tables of performance numbers, and no mention of training/validation splits or generalization across surfaces. Without these data it is impossible to verify that PI-LNO supports the accuracy claims or outperforms the baselines in a load-bearing way.

    Authors: We agree that the benchmark results require clearer quantitative presentation to allow verification. Although the abstract references the studies against UNet, UNet-AM, DeepONet, PI-UNet, and LNO, specific error metrics, tables, data splits, and generalization details were not included in the main text. We will revise the manuscript to add a dedicated results table reporting L2 errors, relative errors, and other metrics for all methods across the contact angle range [20,160], explicitly describe the training/validation/test splits, and include generalization performance on held-out surfaces. revision: yes

  3. Referee: [Training and loss description] Training and loss description: The physics-regularized composite loss is described as combining data fidelity with Navier-Stokes, Cahn-Hilliard, and causality constraints, but no explicit weighting schedule, enforcement mechanism, or ablation study is provided to demonstrate that the physics terms enforce consistency independently of the CFD training data. This leaves open the possibility that outputs reduce to data-driven fits rather than independent physics-informed predictions.

    Authors: We thank the referee for highlighting this gap in the loss description. The composite loss combines the data terms (MSE, MAE, RMSE) with the physics residuals from the Navier-Stokes and Cahn-Hilliard equations plus a causality penalty, but we agree that explicit weights, enforcement details, and an ablation are needed to demonstrate the physics contribution. We will revise the training section to specify the weighting schedule (e.g., data fidelity weight of 1.0 and physics terms at 0.05–0.1), clarify enforcement via automatic differentiation on the predicted fields, and add an ablation study comparing the full PI-LNO to a data-only variant. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The provided manuscript text introduces PI-LNO with the Laplace integral transform as a learned functional basis, trained on external multi-surface CFD data using a composite loss that includes data terms (MSE/MAE/RMSE) plus Navier-Stokes, Cahn-Hilliard, and causality constraints. The central claim that complex Laplace transforms 'natively model the exponential transient dynamics' is asserted as an architectural property but is not accompanied by any quoted equations or steps that reduce the claim to a fitted input, self-definition, or self-citation chain by construction. No uniqueness theorems, ansatzes smuggled via prior self-work, or renamings of known results are exhibited. The approach is benchmarked against independent baselines (UNet, DeepONet, etc.), and the physics constraints serve as regularization rather than forcing the output to equal the input data term. The derivation remains self-contained against external CFD benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The approach rests on standard fluid dynamics equations plus one new modeling choice for the functional basis. Training data range is treated as a free parameter. No independent evidence is given for the invented functional basis beyond the training process.

free parameters (1)
  • Contact angle training range = [20,160]
    Data spans θ_s ∈ [20,160] chosen to cover complex surfaces; this range is selected rather than derived.
axioms (1)
  • domain assumption Navier-Stokes and Cahn-Hilliard equations plus causality constraints govern droplet spreading dynamics
    Invoked directly in the physics-regularized composite loss function.
invented entities (1)
  • Laplace integral transform function as learned physics-informed functional basis no independent evidence
    purpose: To natively model the exponential transient dynamics of droplet spreading within the neural operator
    Core architectural innovation introduced to replace standard bases in neural operators.

pith-pipeline@v0.9.0 · 5518 in / 1427 out tokens · 135760 ms · 2026-05-10T01:02:15.440672+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

42 extracted references · 36 canonical work pages · 2 internal anchors

  1. [1]

    Chakraborty, Microfluidics and Microscale Transport Processes

    S. Chakraborty, Microfluidics and Microscale Transport Processes. CRC Press, 2012

    2012

  2. [2]

    REVIEW PDMS microfluidics: A mini review,

    R. M. Kiran and S. Chakraborty, “REVIEW PDMS microfluidics: A mini review,” 2020, doi: 10.1002/app.48958

    work page doi:10.1002/app.48958 2020

  3. [3]

    Surface Roughness Analysis of Microchannels Featuring Microfluidic Devices Fabricated by Three Different Materials and Methods,

    J. M. Acosta-Cuevas, M. A. García -Ramírez, G. Hinojosa-Ventura, Á. J. Martínez -Gómez, V. H. Pérez -Luna, and O. González -Reynoso, “Surface Roughness Analysis of Microchannels Featuring Microfluidic Devices Fabricated by Three Different Materials and Methods,” Coatings, vol. 13, no. 10, p. 1676, Sep. 2023, doi: 10.3390/coatings13101676

    work page doi:10.3390/coatings13101676 2023

  4. [4]

    Effect of wall roughness on performance of microchannel applied in microfluidic device,

    J. Jia, Q. Song, Z. Liu, and B. Wan g, “Effect of wall roughness on performance of microchannel applied in microfluidic device,” Microsyst. Technol., vol. 25, no. 6, pp. 2385– 2397, Jun. 2019, doi: 10.1007/s00542-018-4124-7

    work page doi:10.1007/s00542-018-4124-7 2019

  5. [5]

    Roughness Induced Boundary Slip in Microchannel Flows,

    C. Kunert and J. Harting, “Roughness Induced Boundary Slip in Microchannel Flows,” Phys. Rev. Lett., vol. 99, no. 17, p. 176001, Oct. 2007, doi: 10.1103/PhysRevLett.99.176001

    work page doi:10.1103/physrevlett.99.176001 2007

  6. [6]

    Influence of surface roughness on the fluid flow in microchannel,

    Y. Li, Z. Zhang, Y. Ji, L. Wang, and D. Li, “Influence of surface roughness on the fluid flow in microchannel,” J. Phys. Conf. Ser. , vol. 27 40, no. 1, p. 012059, Apr. 2024, doi: 10.1088/1742-6596/2740/1/012059

    work page doi:10.1088/1742-6596/2740/1/012059 2024

  7. [7]

    Wetting and Roughness,

    D. Quéré, “Wetting and Roughness,” Annu. Rev. Mater. Res., vol. 38, no. Volume 38, 2008, pp. 71–99, Aug. 2008, doi: 10.1146/annurev.matsci.38.060407.132434

    work page doi:10.1146/annurev.matsci.38.060407.132434 2008

  8. [8]

    Surface Roughness - Hydrophobicity Coupling in Microchannel and Nanochannel Flows,

    M. Sbragaglia, R. Be nzi, L. Biferale, S. Succi, and F. Toschi, “Surface Roughness - Hydrophobicity Coupling in Microchannel and Nanochannel Flows,” Phys. Rev. Lett. , vol. 97, no. 20, p. 204503, Nov. 2006, doi: 10.1103/PhysRevLett.97.204503

    work page doi:10.1103/physrevlett.97.204503 2006

  9. [9]

    Simra: Using crowdsourcing to identify near miss hotspots in bicycle traffic,

    J. Huang, L. J. Segura, T. Wang, G. Zhao, H. Sun, and C. Zhou, “Unsupervised learning for the droplet evolution prediction and process dynamics understanding in inkjet printing,” Addit. Manuf., vol. 35, p. 101197, Oct. 2020, doi: 10.1016/j.addma.2020.101197

    work page doi:10.1016/j.addma.2020.101197 2020

  10. [10]

    Detached Eddy Simulation (DES) of Co-Current Spray Dryer,

    Muhammad Noor Intan Shafinas, M. N. I. Shafinas, Jolius Gimbun, and G. Jolius, “Detached Eddy Simulation (DES) of Co-Current Spray Dryer,” Jan. 2013

    2013

  11. [11]

    Development and future of droplet microfluidics,

    L. Nan, H. Zhang, D. A. Weitz, and H. Cheung Shum, “Development and future of droplet microfluidics,” Lab. Chip, vol. 24, no. 5, pp. 1135–1153, 2024, doi: 10.1039/D3LC00729D

    work page doi:10.1039/d3lc00729d 2024

  12. [12]

    Superior droplet bouncing, anti-icing/anti- frosting and self-cleaning performance of an outstanding superhydrophobic PTFE coating,

    X. Wang, B.-B. Wang, B. Deng, and Z.-M. Xu, “Superior droplet bouncing, anti-icing/anti- frosting and self-cleaning performance of an outstanding superhydrophobic PTFE coating,” Cold Reg. Sci. Technol. , vol. 224, p. 104229, Aug. 2024, doi: 10.1016/j.coldregions.2024.104229

    work page doi:10.1016/j.coldregions.2024.104229 2024

  13. [13]

    Design of Anti-Icing Coatings Using Supercooled Droplets As Nano-to-Microscale Probes | Langmuir

    “Design of Anti-Icing Coatings Using Supercooled Droplets As Nano-to-Microscale Probes | Langmuir.” Accessed: Apr. 07, 2026. [Online]. Available: https://pubs.acs.org/doi/10.1021/la2034565

    work page doi:10.1021/la2034565 2026

  14. [14]

    Atomization characteristics and instabilities in the combustion of multi -component fuel droplets with high volatility differential | Scientific Reports

    “Atomization characteristics and instabilities in the combustion of multi -component fuel droplets with high volatility differential | Scientific Reports.” Accessed: Apr. 07, 2026. [Online]. Available: https://www.nature.com/articles/s41598-017-09663-7

    2026

  15. [15]

    A novel microfluidic device with parallel channels for sperm separation using spermatozoa intrinsic behaviors,

    A. Heydari, M. Zabetian Targhi, I. Halvaei, and R. Nosrati, “A novel microfluidic device with parallel channels for sperm separation using spermatozoa intrinsic behaviors,” Sci. Rep., vol. 13, no. 1, p. 1185, Jan. 2023, doi: 10.1038/s41598-023-28315-7

    work page doi:10.1038/s41598-023-28315-7 2023

  16. [16]

    Recent advances of droplet -based microfluidics for engineering artificial cells,

    S. Fasciano and S. Wang, “Recent advances of droplet -based microfluidics for engineering artificial cells,” SLAS Technol. , vol. 29, no. 2, p. 100090, Apr. 2024, doi: 10.1016/j.slast.2023.05.002

    work page doi:10.1016/j.slast.2023.05.002 2024

  17. [17]

    Arjmandi-Tash, N.M

    O. Arjmandi-Tash, N. M. Kovalchuk, A. Trybala, I. V. Kuchin, and V. Starov, “Kinetics of Wetting and Spreading of Droplets over Various Substrates,” Langmuir, vol. 33, no. 18, pp. 4367–4385, May 2017, doi: 10.1021/acs.langmuir.6b04094

    work page doi:10.1021/acs.langmuir.6b04094 2017

  18. [18]

    Interp lay of Drop Shedding Mechanisms on High Wettability Contrast Biphilic Stripe-Patterned Surfaces,

    H. Tran, Z. He, P. Pirdavari, and M. Y. Pack, “Interp lay of Drop Shedding Mechanisms on High Wettability Contrast Biphilic Stripe-Patterned Surfaces,” Langmuir, vol. 39, no. 48, pp. 17551–17559, Dec. 2023, doi: 10.1021/acs.langmuir.3c03042

    work page doi:10.1021/acs.langmuir.3c03042 2023

  19. [19]

    How can the geo metry of a rough surface affect its wettability? - A coarse-grained simulation analysis,

    S. K. Sethi, R. Gogoi, A. Verma, and G. Manik, “How can the geo metry of a rough surface affect its wettability? - A coarse-grained simulation analysis,” Prog. Org. Coat., vol. 172, p. 107062, Nov. 2022, doi: 10.1016/j.porgcoat.2022.107062

    work page doi:10.1016/j.porgcoat.2022.107062 2022

  20. [20]

    Wettability of porous surfaces,

    A. B. D. Cassie and S. Baxter, “Wettability of porous surfaces,” Trans. Faraday Soc., vol. 40, no. 0, pp. 546–551, Jan. 1944, doi: 10.1039/TF9444000546

    work page doi:10.1039/tf9444000546 1944

  21. [21]

    Two types of Cassie -to-Wenzel wetting transitions on superhydrophobic surfaces during drop impact,

    C. Lee et al. , “Two types of Cassie -to-Wenzel wetting transitions on superhydrophobic surfaces during drop impact,” Soft Matter, vol. 11, no. 23, pp. 4592–4599, 2015

    2015

  22. [22]

    Ezzatneshan, A

    E. Ezzatneshan and A. Khosroabadi, “Droplet spreading dynamics on hydrophobic textured surfaces: A lattice Boltzmann study,” Comput. Fluids, vol. 231, p. 105063, Dec. 2021, doi: 10.1016/j.compfluid.2021.105063

    work page doi:10.1016/j.compfluid.2021.105063 2021

  23. [23]

    Injection continuous liquid interface production of 3D objects,

    G. Lipkowitz et al., “Injection continuous liquid interface production of 3D objects,” Sci. Adv., vol. 8, no. 39, p. eabq3917, Sep. 2022, doi: 10.1126/sciadv.abq3917

    work page doi:10.1126/sciadv.abq3917 2022

  24. [24]

    An Experimental Method for Three-Dimensional Dynamic Contact Angle Analysis,

    D. Baptista, L. Muszyński, D. J. Gardner, and E. Atzema, “An Experimental Method for Three-Dimensional Dynamic Contact Angle Analysis,” J. Adhes. Sci. Technol., vol. 26, no. 18–19, pp. 2199–2215, Oct. 2012, doi: 10.1163/156856111X610135

    work page doi:10.1163/156856111x610135 2012

  25. [25]

    Study on contact spots of fractal rough surfaces based on three -dimensional Weierstrass -Mandelbrot function,

    J. Chen, F. Yang, K. Luo, Y. Wu, C. Niu, and M. Rong, “Study on contact spots of fractal rough surfaces based on three-dimensional Weierstrass-Mandelbrot function,” in 2016 IEEE 62nd Holm Conference on Electrical Contacts (Holm) , Oct. 2016, pp. 198 –204. doi: 10.1109/HOLM.2016.7780032

    work page doi:10.1109/holm.2016.7780032 2016

  26. [26]

    Physics-aware neural operator for high-fidelity fluid dynamics modeling with geometric and spectral priors,

    X. Hu, Q. Ma, P. Zhao, and X. Wang, “Physics-aware neural operator for high-fidelity fluid dynamics modeling with geometric and spectral priors,” Phys. Fluids , vol. 37, no. 11, p. 115111, Nov. 2025, doi: 10.1063/5.0299765

    work page doi:10.1063/5.0299765 2025

  27. [27]

    Sequential Deep Operator Networks (S -DeepONet) for predicting full -field solutions under time -dependent loads,

    J. He, S. Kushwaha, J. Park, S. Koric, D. Abueidda, and I. Jasiuk, “Sequential Deep Operator Networks (S -DeepONet) for predicting full -field solutions under time -dependent loads,” Eng. Appl. Artif. Intell., vol. 127, p. 107258, Jan. 2024, doi: 10.1016/j.engappai.2023.107258

    work page doi:10.1016/j.engappai.2023.107258 2024

  28. [28]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Z. Li et al., “Fourier Neural Operator for Parametric Partial Differential Equations,” May 17, 2021, arXiv: arXiv:2010.08895. doi: 10.48550/arXiv.2010.08895

    work page internal anchor Pith review doi:10.48550/arxiv.2010.08895 2021

  29. [29]

    Prediction of turbulent channel flow using Fourier neural operator-based machine-learning strategy,

    Y. Wang, Z. Li, Z. Yuan, W. Peng, T. Liu, and J. Wang, “Prediction of turbulent channel flow using Fourier neural operator-based machine-learning strategy,” Phys. Rev. Fluids, vol. 9, no. 8, p. 084604, Aug. 2024, doi: 10.1103/PhysRevFluids.9.084604

    work page doi:10.1103/physrevfluids.9.084604 2024

  30. [30]

    Confidence in Assurance 2.0 Cases

    “UNet++: A Nested U -Net Architecture for Medical Image Segmentation | SpringerLink.” Accessed: Jul. 25, 2024. [Online]. Available: https://link.springer.com/chapter/10.1007/978- 3-030-00889-5_1

    work page doi:10.1007/978- 2024

  31. [31]

    Fast and high -precision compressible flowfield inference method of transonic airfoils based on attention UNet,

    Z. ( 朱智杰) Zhu, G. ( 赵国庆) Zhao, and Q. ( 招启军) Zhao, “Fast and high -precision compressible flowfield inference method of transonic airfoils based on attention UNet,” Phys. Fluids, vol. 36, no. 3, p. 036111, Mar. 2024, doi: 10.1063/5.0188550

    work page doi:10.1063/5.0188550 2024

  32. [32]

    Lattice-Boltzmann-Driven Physics-Informed Neural Networks for Droplet Wettability on Rough Surfaces

    G. S. Meshram, P. P. Chakrabarti, and S. Chakraborty, “Lattice-Boltzmann-Driven Physics- Informed Neural Networks for Droplet Wettability on Rough Surfaces,” Apr. 03, 2026, arXiv: arXiv:2604.03481. doi: 10.48550/arXiv.2604.03481

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.03481 2026

  33. [33]

    Operator Learning: A Statistical Perspective

    U. Subedi and A. Tewari, “Operator Learning: A Statistical Perspective”

  34. [34]

    Fourier neural operator with boundary conditions for efficient prediction of steady airfoil flows,

    Y. Dai, Y. An, Z. Li, J. Zhang, and C. Yu, “Fourier neural operator with boundary conditions for efficient prediction of steady airfoil flows,” Appl. Math. Mech., vol. 44, no. 11, pp. 2019– 2038, Nov. 2023, doi: 10.1007/s10483-023-3050-9

    work page doi:10.1007/s10483-023-3050-9 2019

  35. [35]

    An attention -enhanced Fourier neural operator model for predicting flow fields in turbomachinery Cascades,

    L. Li, W. Zhang, Y. Li, C. Jiang, and Y. Wang, “An attention -enhanced Fourier neural operator model for predicting flow fields in turbomachinery Cascades,” Phys. Fluids, vol. 37, no. 3, p. 036121, Mar. 2025, doi: 10.1063/5.0254681

    work page doi:10.1063/5.0254681 2025

  36. [36]

    Fourier neural operator for high -resolution fluid flow simulation based on low -resolution data: the vorticity equation as an example,

    H. Qu, X. Zheng, L. Yang, and Z. Song, “Fourier neural operator for high -resolution fluid flow simulation based on low -resolution data: the vorticity equation as an example,” Acta Oceanol. Sin., vol. 44, no. 6, pp. 165–177, Jun. 2025, doi: 10.1007/s13131-024-2453-1

    work page doi:10.1007/s13131-024-2453-1 2025

  37. [37]

    Learning nonlinear operators in latent spaces for real-time predictions of complex dynamics in physical systems,

    K. Kontolati, S. Goswami, G. Em Karniadakis, and M. D. Shields, “Learning nonlinear operators in latent spaces for real-time predictions of complex dynamics in physical systems,” Nat. Commun., vol. 15, no. 1, p. 5101, Jun. 2024, doi: 10.1038/s41467-024-49411-w

    work page doi:10.1038/s41467-024-49411-w 2024

  38. [38]

    Operator Learning for Reconstructing Flow Fields from Sparse Measurements:

    Q. Zhang, D. Krotov, and G. E. Karniadakis, “Operator learning for reconstructing flow fields from sparse measurements: An energy transformer approach,” J. Comput. Phys., vol. 538, p. 114148, Oct. 2025, doi: 10.1016/j.jcp.2025.114148

    work page doi:10.1016/j.jcp.2025.114148 2025

  39. [39]

    Enhancing generalization in endwall film cooling prediction: Incorporating the superposition principle into transformer -based neural operators,

    Q. Wang, L. Song, T. Liu, and Z. Guo, “Enhancing generalization in endwall film cooling prediction: Incorporating the superposition principle into transformer -based neural operators,” Phys. Fluids, vol. 36, no. 12, p. 126110, Dec. 2024, doi: 10.1063/5.0239483

    work page doi:10.1063/5.0239483 2024

  40. [40]

    A physics-informed transformer neural operator for learning generalized solutions of initial boundary value problems,

    S. K. Boya and D. Subramani, “A physics-informed transformer neural operator for learning generalized solutions of initial boundary value problems,” May 14, 2025, arXiv: arXiv:2412.09009. doi: 10.48550/arXiv.2412.09009

    work page doi:10.48550/arxiv.2412.09009 2025

  41. [41]

    Scaling the predictions of multiphase flow through porous media using operator learning,

    N. Jain, S. Roy, H. Kodamana, and P. Nair, “Scaling the predictions of multiphase flow through porous media using operator learning,” Chem. Eng. J., vol. 503, p. 157671, Jan. 2025, doi: 10.1016/j.cej.2024.157671

    work page doi:10.1016/j.cej.2024.157671 2025

  42. [42]

    Géron, Hands-On Machine Learning with Scikit-Learn , Keras , and

    A. Géron, Hands-On Machine Learning with Scikit-Learn , Keras , and. O’Reilly Media