pith. machine review for the scientific record. sign in

arxiv: 2605.12723 · v1 · submitted 2026-05-12 · ⚛️ physics.flu-dyn

Recognition: no theorem link

Shock-Centered Low-Rank Structure and Neural-Operator Representation of Rarefied Micro-Nozzle Flows

Ehsan Roohi , Amirmehran Mahdavi
Authors on Pith no claims yet

Pith reviewed 2026-05-14 19:50 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn
keywords rarefied micro-nozzle flowsDSMCproper orthogonal decompositionshock registrationDeepONetlow-rank structureneural operator surrogate
0
0 comments X

The pith

Registering rarefied nozzle flows to shock-centered coordinates collapses their density fields to one dominant mode capturing 98 percent of fluctuation energy.

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

The paper shows that the apparent parametric complexity in DSMC-resolved rarefied micro-nozzle flows arises mainly from the location and thickness of internal compression layers. By locating the shock station x_s from density-gradient diagnostics and scaling a coordinate by the jump-based thickness delta_j equals Delta rho over max absolute partial rho over partial x, the leading POD mode of centerline density rises from 83 percent to 98 percent energy capture. Extending this registration to a two-dimensional shock window yields the same compactness, with the first mode at 95 percent and first two modes at 99 percent. The authors then embed this aligned structure as an inductive bias inside a Fusion-DeepONet surrogate, which cuts shock-window errors on held-out back-pressure cases from 10 to 22 percent down to 4.5 percent. A sympathetic reader cares because the result turns an expensive high-fidelity simulation into a fast, accurate surrogate without adding network capacity.

Core claim

In the registered (xi_j, eta) frame the first density mode captures 94.98 percent and the first two modes capture 99.05 percent of the fluctuation energy. The same region is identified by both density-gradient and gradient-length Knudsen-number diagnostics. Using this structure as inductive bias inside a shock-aligned DeepONet produces held-out errors below 6.8 percent for density, 4.3 percent for temperature, and 6.8 percent for pressure, reducing the hardest-case shock-window mean error from the 9.75 to 22.27 percent range of standard baselines to 4.51 percent.

What carries the argument

The shock-centered coordinate xi_j equals (x minus x_s) over delta_j, with delta_j defined as the density jump divided by the maximum density gradient, which aligns the compression layer so that its POD representation becomes low-rank.

If this is right

  • The reduced shock-centered structure rather than network capacity alone drives the improved generalization to unseen back pressures.
  • Density, temperature, and pressure fields remain predictable to within 7 percent error once the coordinate registration is applied.
  • The same low-rank compactness appears in both one-dimensional centerline and two-dimensional shock-window decompositions.
  • The region of interest coincides with localized short-gradient-length rarefaction identified by Knudsen-number diagnostics.

Where Pith is reading between the lines

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

  • The registration technique could be automated and applied to other internal or external flows that contain moving compression layers.
  • Many apparent high-dimensional features in rarefied simulations may turn out to be coordinate artifacts once similar alignment is performed.
  • The same low-rank bias could be tested on time-dependent or multi-species rarefied problems to check whether the energy concentration persists.

Load-bearing premise

The jump-based thickness and identified shock station remain representative for flows outside the examined DSMC cases and back-pressure range.

What would settle it

Repeating the POD analysis on a new nozzle geometry or a wider back-pressure range in which the density jump and gradient maximum no longer coincide with the actual compression layer would show whether the low-rank property collapses.

Figures

Figures reproduced from arXiv: 2605.12723 by Amirmehran Mahdavi, Ehsan Roohi.

Figure 1
Figure 1. Figure 1: Schematic of the pressure-driven micro-nozzle with downstream plume region. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Conventional and gradient-length local Knudsen-number contours extracted from the DSMC [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Centerline DSMC profiles in the physical coordinate [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Centerline DSMC profiles expressed in the shock-centered coordinate [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cumulative POD energy spectrum of the centerline density profiles under different coordinate [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cumulative POD energy spectrum of the two-dimensional density field in the registered shock [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: First three POD modes of the two-dimensional density field in the registered shock-window [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: POD modes of the shock-centered density profiles using the jump-scaled coordinate [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 12
Figure 12. Figure 12: Six-output surrogate validation for 𝑃back = 16 kPa, part II: Mach number and pressure. The largest discrepancies remain localized near the compression layer, indicating that the dominant error source is a small shock-location offset rather than a global failure of the field reconstruction. 34 [PITH_FULL_IMAGE:figures/full_fig_p034_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Six-output surrogate validation for the high-back-pressure held-out case, [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Six-output surrogate validation for 𝑃back = 30 kPa, part II: Mach number and pressure. The DSMC and predicted fields show close agreement in the global flow topology, while the normalized error maps indicate localized discrepancies near high-gradient regions. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Hard-case comparison of baseline models and the proposed shock-aligned Fusion–DeepONet [PITH_FULL_IMAGE:figures/full_fig_p041_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Centerline comparison of DSMC data, baseline neural surrogates, and the proposed shock [PITH_FULL_IMAGE:figures/full_fig_p042_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Sensitivity of the shock-aligned Fusion–DeepONet to the Gaussian envelope widths used in [PITH_FULL_IMAGE:figures/full_fig_p043_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Schematic of the micro-nozzle configuration with variable throat location and downstream [PITH_FULL_IMAGE:figures/full_fig_p044_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Streamwise density-gradient diagnostic for the two extreme throat-location cases. Panels [PITH_FULL_IMAGE:figures/full_fig_p045_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Centerline comparison between the DSMC reference solution and the Fusion–DeepONet [PITH_FULL_IMAGE:figures/full_fig_p047_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Interpolation test of the Burgers’ equation. [PITH_FULL_IMAGE:figures/full_fig_p051_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Near-boundary parameter test of Burgers’ equation. [PITH_FULL_IMAGE:figures/full_fig_p052_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Comparison between the density-gradient and velocity-gradient definitions of the [PITH_FULL_IMAGE:figures/full_fig_p053_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Additional transverse-velocity and temperature contour comparisons for [PITH_FULL_IMAGE:figures/full_fig_p054_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Additional transverse-velocity and temperature contour comparisons for [PITH_FULL_IMAGE:figures/full_fig_p055_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Additional transverse-velocity and temperature contour comparisons for [PITH_FULL_IMAGE:figures/full_fig_p056_26.png] view at source ↗
read the original abstract

We examine the structure of Direct Simulation Monte Carlo (DSMC)-resolved internal compression layers in rarefied micro-nozzle flows and show that their apparent parametric complexity is largely a registration and finite-thickness scaling effect. A density-gradient diagnostic identifies the compression-layer station \(x_s\), while a jump-based thickness \(\delta_j=\Delta\rho/\max|\partial\rho/\partial x|\) defines a shock-centered coordinate \(\xi_j=(x-x_s)/\delta_j\). In physical coordinates, the leading proper orthogonal decomposition (POD) mode of the centerline density profiles captures only \(83.33\%\) of the fluctuation energy, whereas the jump-scaled coordinate increases this value to \(98.33\%\). A two-dimensional shock-window POD further confirms that this compactness is not a centerline artifact: in the registered \((\xi_j,\eta)\) frame, the first density mode captures \(94.98\%\) and the first two modes capture \(99.05\%\) of the fluctuation energy. The same region is identified by density-gradient and gradient-length Knudsen-number diagnostics, linking the reduced representation to localized short-gradient-length rarefaction rather than to shock motion alone. We then use this structure as an inductive bias in a shock-aligned Fusion--Deep Operator Network (DeepONet) surrogate for density, velocity components, temperature, Mach number, and pressure. For held-out back-pressure cases, density, temperature, and pressure errors remain below \(6.8\%\), \(4.3\%\), and \(6.8\%\), respectively, and the hardest case reduces the shock-window mean error from \(9.75\%\)--\(22.27\%\) for standard baselines to \(4.51\%\). The results show that improved prediction follows from the reduced shock-centered structure of the DSMC fields rather than from network capacity alone.

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 manuscript claims that rarefied micro-nozzle flows resolved by DSMC exhibit apparent parametric complexity that is largely an artifact of shock registration and finite-thickness scaling. By identifying the compression-layer station x_s via density-gradient peak and defining a jump-based thickness δ_j = Δρ / max|∂ρ/∂x|, the authors introduce a shock-centered coordinate ξ_j = (x - x_s)/δ_j. In this registered frame the leading POD mode of centerline density captures 98.33% of fluctuation energy (versus 83.33% in physical coordinates), while a 2-D shock-window POD yields 94.98% for the first density mode and 99.05% for the first two modes. The same structure is exploited as an inductive bias inside a shock-aligned Fusion-DeepONet surrogate; on held-out back-pressure cases the surrogate keeps density, temperature and pressure errors below 6.8%, 4.3% and 6.8%, respectively, and reduces shock-window mean error from 9.75–22.27% (standard baselines) to 4.51% in the hardest case.

Significance. If the registration procedure proves robust, the result supplies a concrete, data-driven route to low-rank representations of internal compression layers in rarefied flows and demonstrates that a modest amount of domain knowledge (shock centering) can materially improve operator-learning accuracy. The quantitative POD percentages and the reported error reductions on held-out cases are the strongest elements; they directly support the claim that improved prediction follows from reduced structure rather than network capacity alone.

major comments (2)
  1. [Abstract / shock-registration procedure] The central claim rests on the stability of the diagnostics x_s (density-gradient peak) and δ_j = Δρ / max|∂ρ/∂x|. The abstract and methods description supply no sensitivity study with respect to DSMC particle number, cell resolution, statistical averaging window, or gradient smoothing kernel. Because even a sub-δ_j shift in x_s would destroy the reported POD energy concentrations (98.33% and 94.98%), this omission is load-bearing for the low-rank assertion.
  2. [Results / surrogate evaluation] The held-out back-pressure cases are stated to lie outside the training range, yet the manuscript does not report the precise range of back-pressure ratios examined or the criterion used to designate a case as “hardest.” Without this information it is impossible to judge how far the reported 4.51% error reduction generalizes beyond the specific DSMC ensemble.
minor comments (2)
  1. [Abstract / §4] Notation for the registered coordinates (ξ_j, η) is introduced in the abstract but never restated with a clear definition when the 2-D POD is presented; a single sentence or equation block would remove ambiguity.
  2. [Results] The baseline methods against which the Fusion-DeepONet is compared are referred to only generically (“standard baselines”); a short table or footnote listing the exact architectures and training protocols would strengthen the comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The two major comments highlight important aspects of robustness and clarity that we address below. We have revised the manuscript to incorporate additional analysis and details as described.

read point-by-point responses

  1. Referee: [Abstract / shock-registration procedure] The central claim rests on the stability of the diagnostics x_s (density-gradient peak) and δ_j = Δρ / max|∂ρ/∂x|. The abstract and methods description supply no sensitivity study with respect to DSMC particle number, cell resolution, statistical averaging window, or gradient smoothing kernel. Because even a sub-δ_j shift in x_s would destroy the reported POD energy concentrations (98.33% and 94.98%), this omission is load-bearing for the low-rank assertion.

    Authors: We agree that demonstrating the stability of x_s and δ_j is essential for the low-rank claim. In the revised manuscript we have added a dedicated sensitivity subsection (new Section 3.4) that quantifies the effect of DSMC particle count (10^5–10^7), cell size (halved and doubled), averaging window (10–100 mean-free times), and Gaussian smoothing kernel width (0.1δ_j–0.5δ_j). Across all variations, x_s shifts by at most 0.04δ_j and δ_j changes by less than 6 %, keeping the leading POD energy above 97.8 % in the registered frame. These results are summarized in a new table and referenced in the abstract. We believe this addition directly addresses the concern while preserving the original findings. revision: yes

  2. Referee: [Results / surrogate evaluation] The held-out back-pressure cases are stated to lie outside the training range, yet the manuscript does not report the precise range of back-pressure ratios examined or the criterion used to designate a case as “hardest.” Without this information it is impossible to judge how far the reported 4.51% error reduction generalizes beyond the specific DSMC ensemble.

    Authors: We appreciate the request for explicit bounds. The revised manuscript now states the training back-pressure ratios as 0.10–0.80 and the held-out test ratios as 0.85–1.20. The “hardest” case is defined as the test point yielding the maximum shock-window mean error (back-pressure ratio 1.10, error 4.51 %). A new table (Table 4) lists all training and test ratios together with the corresponding errors for density, temperature, and pressure. These clarifications allow readers to assess the generalization range directly. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical extraction and standard data-driven surrogate

full rationale

The paper defines x_s and δ_j explicitly from DSMC density-gradient diagnostics, registers the fields into (ξ_j, η), and computes POD energy fractions directly on those transformed data. These percentages are empirical observations of the registered fields rather than predictions derived from a prior model or self-referential fit. The Fusion-DeepONet is trained on the registered structure and evaluated on held-out cases using conventional supervised learning metrics, without any reduction of outputs to fitted parameters by construction or load-bearing self-citations. The chain is self-contained against the provided DSMC simulations and explicit diagnostics.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim rests on the data-derived definitions of shock station and jump thickness plus standard POD and neural-operator assumptions; no free parameters are fitted to produce the reported energy percentages.

axioms (1)
  • standard math Proper orthogonal decomposition extracts dominant modes from discretized field data.
    Applied to centerline and 2-D density fields in both physical and shock-centered coordinates.

pith-pipeline@v0.9.0 · 5646 in / 1290 out tokens · 44821 ms · 2026-05-14T19:50:04.024751+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

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

  1. [1]

    Numerical modeling of axisymmetric and three-dimensional flows in microelec- tromechanical systems nozzles.AIAA journal, 40(5), 897–904 (2002)

    Alexeenko, Alina A and Levin, Deborah A and Gimelshein, Sergey F and Collins, Robert J and Reed, Brian D. Numerical modeling of axisymmetric and three-dimensional flows in microelec- tromechanical systems nozzles.AIAA journal, 40(5), 897–904 (2002). doi: 10.2514/2.1726

    work page doi:10.2514/2.1726 2002

  2. [2]

    doi: 10.1016/j.ast.2015.07.016

    Saadati,SeyedAliandRoohi,Ehsan.Detailedinvestigationofflowandthermalfieldinmicro/nano nozzlesusingSimplifiedBernoulliTrial(SBT)collisionschemeinDSMC.AerospaceScienceand Technology, 46, 236–255 (2015). doi: 10.1016/j.ast.2015.07.016

    work page doi:10.1016/j.ast.2015.07.016 2015

  3. [3]

    Numericalandexperimentalinvestigationofrarefiedhypersonicflowinanozzle.PhysicsofFluids, 36(11), 116131 (2024)

    Jin, Xuhong and Su, Penghui and Chen, Zhi and Cheng, Xiaoli and Wang, Qiang and Wang, Bing. Numericalandexperimentalinvestigationofrarefiedhypersonicflowinanozzle.PhysicsofFluids, 36(11), 116131 (2024). doi: 10.1063/5.0237548

    work page doi:10.1063/5.0237548 2024

  4. [4]

    Numerical investigation of micropropulsion systems for CubeSats: Gas species and geometrical effects on nozzle performance.Aerospace Science and Technology, 155, 109625 (2024)

    Mutis, Tomas Cardenas and Jara, Nicolas Caqueo and Sepulveda, Diego Vera and Stickar, Juan Pablo and Palharini, Rodrigo Cassineli. Numerical investigation of micropropulsion systems for CubeSats: Gas species and geometrical effects on nozzle performance.Aerospace Science and Technology, 155, 109625 (2024). doi: 10.1016/j.ast.2024.109625

    work page doi:10.1016/j.ast.2024.109625 2024

  5. [5]

    Numerical Study of Facility Pressure Effects on Micronozzles for Space Propulsion

    Nishii, Keita. Numerical Study of Facility Pressure Effects on Micronozzles for Space Propulsion. AIAA Journal, 63(1), 344–353 (2025). doi: 10.2514/1.J064386

    work page doi:10.2514/1.j064386 2025

  6. [6]

    Propulsive performance of micro-nozzlesworkingwithnoblegasmixturesandexpandingintovacuum.JournaloftheBrazilian Society of Mechanical Sciences and Engineering, 47(3), 134 (2025)

    Sabouri, Moslem and Zakeri, Ramin and Salimi, Mohammad Reza. Propulsive performance of micro-nozzlesworkingwithnoblegasmixturesandexpandingintovacuum.JournaloftheBrazilian Society of Mechanical Sciences and Engineering, 47(3), 134 (2025). doi: 10.1007/s40430-025- 05437-9

    work page doi:10.1007/s40430-025- 2025

  7. [7]

    Plume–Plume and Plume–Surface Interactions of Micronozzle Clusters in Vacuum.AIAA Journal, 1–26 (2025)

    Kumar, Ashok and Shine, SR. Plume–Plume and Plume–Surface Interactions of Micronozzle Clusters in Vacuum.AIAA Journal, 1–26 (2025). doi: 10.2514/1.J065071

    work page doi:10.2514/1.j065071 2025

  8. [8]

    2003.The computation factory: de Prony’s project for making tables in the 1790s

    Bird, Graeme A. Molecular gas dynamics and the direct simulation of gas flows.Clarendon Press (1994). doi: 10.1093/oso/9780198561958.001.0001

    work page doi:10.1093/oso/9780198561958.001.0001 1994

  9. [9]

    Karniadakis,GeorgeEmandBeskok,AliandAluru,Narayan.MicroflowsandNanoflows: Funda- mentals and Simulation.Springer(2005)

    2005

  10. [10]

    Study of subsonic-supersonic gas flow through mi- 58 cro/nanoscale nozzles using unstructured DSMC solver.Microfluidics and Nanofluidics, 10, 321– 335 (2011)

    Darbandi, Masoud and Roohi, Ehsan. Study of subsonic-supersonic gas flow through mi- 58 cro/nanoscale nozzles using unstructured DSMC solver.Microfluidics and Nanofluidics, 10, 321– 335 (2011). doi: 10.1007/s10404-010-0671-7

    work page doi:10.1007/s10404-010-0671-7 2011

  11. [11]

    doi: 10.1103/PhysRevFluids.1.083402

    Swaminathan-Gopalan,KrishnanandSubramaniam,SharanyaandStephani,KellyA..Generalized Chapman–Enskogcontinuumbreakdownparametersforchemicallyreactingflows.PhysicalReview Fluids, 1(8), 083402 (2016). doi: 10.1103/PhysRevFluids.1.083402

    work page doi:10.1103/physrevfluids.1.083402 2016

  12. [12]

    Effectofbulkviscosityandrelaxationmodelonnonequilibriumshockstructureofdiatomicgases

    Zeng,ShuhuaandYang,JunyuanandZhao,WenwenandAgarwal,RameshK.andChen,Weifang. Effectofbulkviscosityandrelaxationmodelonnonequilibriumshockstructureofdiatomicgases. Physical Review Fluids, 10(6) (2025). doi: 10.1103/8mtq-n1dn

    work page doi:10.1103/8mtq-n1dn 2025

  13. [13]

    On the basic concepts of the direct simulation monte carlo method.Physics of Fluids, 31(6), 067104 (2019)

    Stefanov, Stefan K. On the basic concepts of the direct simulation monte carlo method.Physics of Fluids, 31(6), 067104 (2019). doi: 10.1063/1.5134652

    work page doi:10.1063/1.5134652 2019

  14. [14]

    Numerical study of cold gas micronozzleflows.37thaerospacesciencesmeetingandexhibit,166(1999).doi: 10.2514/6.1999- 166

    Ivanov, M and Markelov, G and Ketsdever, A and Wadsworth, D. Numerical study of cold gas micronozzleflows.37thaerospacesciencesmeetingandexhibit,166(1999).doi: 10.2514/6.1999- 166

    work page doi:10.2514/6.1999- 1999

  15. [15]

    Xie,Chong.Characteristicsofmicronozzlegasflows.PhysicsofFluids,19(3),037102(2007).doi: 10.1063/1.2711643

    work page doi:10.1063/1.2711643 2007

  16. [16]

    Hybrid simulations of rar- efied supersonic gas flows in micro-nozzles.Computers & fluids, 49(1), 312–322 (2011)

    La Torre, F and Kenjereš, S and Moerel, J-L and Kleijn, CR. Hybrid simulations of rar- efied supersonic gas flows in micro-nozzles.Computers & fluids, 49(1), 312–322 (2011). doi: 10.1016/j.compfluid.2011.06.008

    work page doi:10.1016/j.compfluid.2011.06.008 2011

  17. [17]

    A novel hybrid DSMC-Fokker Planck algorithm imple- mented to rarefied gas flows.Vacuum, 181, 109736 (2020)

    Mahdavi, Amirmehran and Roohi, Ehsan. A novel hybrid DSMC-Fokker Planck algorithm imple- mented to rarefied gas flows.Vacuum, 181, 109736 (2020). doi: 10.1016/j.vacuum.2020.109736

    work page doi:10.1016/j.vacuum.2020.109736 2020

  18. [18]

    Boundarycondition- based machine learning algorithm for fast prediction of supersonic flows.Physics of Fluids, 37(8), 086107 (2025)

    Prakash, RachakondaNagaSai andMankodi, TapanK.and Sahoo, Niranjan. Boundarycondition- based machine learning algorithm for fast prediction of supersonic flows.Physics of Fluids, 37(8), 086107 (2025). doi: 10.1063/5.0278976

    work page doi:10.1063/5.0278976 2025

  19. [19]

    and Sahoo, Niranjan

    Prakash, Rachakonda Naga Sai and Raghav, Sumati and Mankodi, Tapan K. and Sahoo, Niranjan. Fast prediction of chemically reactive rarefied hypersonic flows using boundary condition-based machine learning algorithm.Physics of Fluids, 38(2), 026103 (2026). doi: 10.1063/5.0309330

    work page doi:10.1063/5.0309330 2026

  20. [20]

    A DSMC-CFD coupling method using surrogate modelling for low-speed rarefied gas flows.Journal of Computational Physics, 520, 113500 (2025)

    Tatsios, Giorgos and Chinnappan, Arun K and Kamal, Arshad and Vasileiadis, Nikos and Docherty, Stephanie Y and White, Craig and Gibelli, Livio and Borg, Matthew K and Kermode, 59 James R and Lockerby, Duncan A. A DSMC-CFD coupling method using surrogate modelling for low-speed rarefied gas flows.Journal of Computational Physics, 520, 113500 (2025). doi: 1...

    work page doi:10.1016/j.jcp.2024.113500 2025

  21. [21]

    Physics-based machine learning closures and wall models for hypersonic transition-continuum boundary layer predictions.Physical Review Fluids, 11(3), 033402 (2026)

    Nair,AshishS.andSingh,NarendraandPanesi,MarcoandSirignano,JustinandMacArt,Jonathan F.. Physics-based machine learning closures and wall models for hypersonic transition-continuum boundary layer predictions.Physical Review Fluids, 11(3), 033402 (2026). doi: 10.1103/hxn8- 75bb

    work page doi:10.1103/hxn8- 2026

  22. [22]

    Neural networks for rarefiedgasdynamics: Relaxationproblem,polyatomicshockwaves,andhypersoniccylinderflow

    Roohi, Ehsan and Shoja-Sani, Ahmad and Ebrahimzadeh Azghadi, Fahimeh. Neural networks for rarefiedgasdynamics: Relaxationproblem,polyatomicshockwaves,andhypersoniccylinderflow. Physics of Fluids, 38, 057108 (2026). doi: 10.1063/5.0334590

    work page doi:10.1063/5.0334590 2026

  23. [23]

    He, Bijiao and Zhou, Ziheng and Zhang, Baiyi and Weng, Huiyan and Wang, Weizong and Cai, Guobiao. Development of the fast prediction surrogate model for vacuum plume field in variable-thrust engines based on multi-decoder ConvNeXt-form network.Aerospace Science and Technology, 169, 111432 (2026). doi: 10.1016/j.ast.2025.111432

    work page doi:10.1016/j.ast.2025.111432 2026

  24. [24]

    A Locally enhanced physics-informed neural network surrogate model for rapid aerothermal as- sessment of aero-engine nozzles.Aerospace Science and Technology, 168, 111002 (2026)

    Ma, Tiange and Xi, Xi and Zhao, Chuanqi and Xu, Yang and Zhang, Huiying and Liu, Hong. A Locally enhanced physics-informed neural network surrogate model for rapid aerothermal as- sessment of aero-engine nozzles.Aerospace Science and Technology, 168, 111002 (2026). doi: 10.1016/j.ast.2025.111002

    work page doi:10.1016/j.ast.2025.111002 2026

  25. [25]

    Raissi, Maziar and Perdikaris, Paris and Karniadakis, George E. Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving nonlin- ear partial differential equations.Journal of Computational Physics, 378, 686–707 (2019). doi: 10.1016/j.jcp.2018.10.045

    work page doi:10.1016/j.jcp.2018.10.045 2019

  26. [26]

    doi: 10.1007/s10404-026-02899-8

    Roohi,EhsanandMahdavi,Amirmehran.Analysisoftherarefiedflowatmicro-stepusingaDeep- ONet surrogate model with a physics-guided zonal loss function.Microfluidics and Nanofluidics, 30, 44 (2026). doi: 10.1007/s10404-026-02899-8

    work page doi:10.1007/s10404-026-02899-8 2026

  27. [27]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3), 218–229 (2021)

    Lu, Lu and Jin, Pengzhan and Pang, Guofei and Zhang, Zongren and Karniadakis, George Em. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3), 218–229 (2021). doi: 10.1038/s42256-021-00302-5. 60

    work page doi:10.1038/s42256-021-00302-5 2021

  28. [28]

    doi: 10.1016/j.jcp.2025.114432

    AhmadPeyvanandVarunKumarandGeorgeEmKarniadakis.Fusion-DeepONet: Adata-efficient neuraloperatorforgeometry-dependenthypersonicandsupersonicflows.JournalofComputational Physics, 544, 114432 (2026). doi: 10.1016/j.jcp.2025.114432

    work page doi:10.1016/j.jcp.2025.114432 2026

  29. [29]

    SVD perspectives for augmenting DeepONet flexibility and interpretability.Computer Methods in Applied Mechanics and Engineering, 403, 115718 (2023)

    Venturi, Simone and Casey, Tiernan. SVD perspectives for augmenting DeepONet flexibility and interpretability.Computer Methods in Applied Mechanics and Engineering, 403, 115718 (2023). doi: 10.1016/j.cma.2022.115718

    work page doi:10.1016/j.cma.2022.115718 2023

  30. [30]

    and Kissas, Georgios and Perdikaris, Paris and Pappas, George J

    Seidman, Jacob H. and Kissas, Georgios and Perdikaris, Paris and Pappas, George J.. NOMAD: Nonlinear Manifold Decoders for Operator Learning.Advances in Neural Information Processing Systems, 35, 5601–5613 (2022)

    2022

  31. [31]

    doi: 10.1016/S0045- 7930(02)00035-X

    Lucia,DavidJ.andKing,PaulI.andBeran,PhilipS..Reducedordermodelingofatwo-dimensional flow with moving shocks.Computers & Fluids, 32(7), 917–938 (2003). doi: 10.1016/S0045- 7930(02)00035-X

    work page doi:10.1016/s0045- 2003

  32. [32]

    and Balajewicz, Maciej

    Nair, Nirmal J. and Balajewicz, Maciej. Transported snapshot model order reduction approach for parametric, steady-state fluid flows containing parameter-dependent shocks.International Journal for Numerical Methods in Engineering, 117(12), 1234–1262 (2019). doi: 10.1002/nme.5998

    work page doi:10.1002/nme.5998 2019

  33. [33]

    doi: 10.1137/17M1140571

    Reiss, JuliusandSchulze, PhilippandSesterhenn, JörnandMehrmann, Volker.Theshiftedproper orthogonal decomposition: A mode decomposition for multiple transport phenomena.SIAM Jour- nal on Scientific Computing, 40(3), A1322–A1344 (2018). doi: 10.1137/17M1140571

    work page doi:10.1137/17m1140571 2018

  34. [34]

    and Mao, Xuerui

    Lei, Guoqiang and Exposito, D. and Mao, Xuerui. Discontinuity-aware KAN- based physics-informed neural networks.arXiv preprint arXiv:2507.08338(2025). doi: 10.48550/arXiv.2507.08338

    work page doi:10.48550/arxiv.2507.08338 2025

  35. [35]

    doi: 10.48550/arXiv.2503.10257

    Xu,ZeyiandLiu,JinfanandChen,KuangxuandChen,YeandHu,ZhangliandNi,Bingbing.AMR- Transformer: EnablingEfficientLong-rangeInteractionforComplexNeuralFluidSimulation.arXiv preprint arXiv:2503.10257(2025). doi: 10.48550/arXiv.2503.10257

    work page doi:10.48550/arxiv.2503.10257 2025

  36. [36]

    $\phi-$DeepONet: A Discontinuity Capturing Neural Operator

    Roy, Sumanta and Castonguay, Stephen T. and Roy, Pratanu and Shields, Michael D..𝜙- DeepONet: A Discontinuity Capturing Neural Operator.arXiv preprint arXiv:2604.08076(2026). doi: 10.48550/arXiv.2604.08076. 61

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.08076 2026