pith. sign in

arxiv: 2604.24225 · v1 · submitted 2026-04-27 · ⚛️ physics.flu-dyn

Multilevel radial basis function surrogates for noise-robust DSMC-CFD coupling

Arshad Kamal , Arun K. Chinnappan , James R. Kermode , Duncan A. Lockerby This is my paper

Pith reviewed 2026-05-08 01:30 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn
keywords hybrid simulationDSMC-CFD couplingradial basis functionsrarefied gas flowsBayesian surrogatesnoise robustnesscomplex geometriesmultilevel methods
0
0 comments X

The pith

Multilevel radial basis functions let the MMS-Sparse method couple DSMC to CFD on complex geometries while staying robust to particle noise.

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

The paper develops a way to make hybrid particle-continuum simulations practical for real engineering geometries. Earlier versions of the MMS-Sparse approach used Bayesian surrogates built on global basis functions to supply smooth corrections to the continuum solver and thereby suppress statistical noise. Those global functions restricted the method to simple domains. The new work replaces them with multilevel radial basis functions that can capture both large-scale and fine-scale features locally. This change preserves the noise-robust and automated character of the Bayesian method while allowing the hybrid scheme to be applied to arbitrary shapes. Validation on a two-dimensional lid-driven cavity flow coupled to an OpenFOAM finite-volume solver shows results that match benchmark solutions.

Core claim

The central claim is that multilevel radial basis functions can be substituted for global polynomials inside the MMS-Sparse framework so that the resulting smooth constitutive corrections remain accurate enough for stable DSMC-CFD coupling on spatially complex domains, as shown by agreement with reference solutions on the lid-driven cavity test case.

What carries the argument

Multilevel radial basis functions that combine coarse and fine local kernels to represent the Bayesian surrogate corrections to the CFD constitutive relations at multiple spatial scales.

If this is right

  • The hybrid solver can be applied to arbitrary two-dimensional domains without requiring geometry-specific manual tuning of basis functions.
  • Statistical noise from the DSMC region continues to be filtered by the Bayesian surrogate, preserving stability of the coupled system.
  • The method integrates directly with existing finite-volume CFD codes such as OpenFOAM.
  • The same surrogate construction retains the automation and noise-robustness properties of the original MMS-Sparse approach.

Where Pith is reading between the lines

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

  • The local character of the multilevel functions suggests the surrogate could be refined adaptively in regions of high flow gradient, potentially lowering overall computational cost.
  • The same replacement of global by multilevel bases might be tested in three-dimensional hybrid simulations of devices with realistic internal geometry.
  • Because the corrections remain smooth, the approach could be examined for other particle-continuum couplings that suffer from statistical fluctuations, such as certain plasma or granular-flow problems.

Load-bearing premise

That multilevel radial basis functions can represent the required smooth constitutive corrections accurately enough across complex geometries without introducing new errors or instabilities into the DSMC-CFD coupling.

What would settle it

A quantitative comparison of hybrid simulation results against a converged full-DSMC reference on a geometry containing sharp corners or curved boundaries, checking whether any macroscopic flow quantities diverge beyond the level of statistical fluctuations.

Figures

Figures reproduced from arXiv: 2604.24225 by Arshad Kamal, Arun K. Chinnappan, Duncan A. Lockerby, James R. Kermode.

Figure 1
Figure 1. Figure 1: The geometry of the 2D lid-driven cavity problem, w view at source ↗
Figure 2
Figure 2. Figure 2: (a) Proof-of-concept of MMS-Sparse. (i) Noisy DSM view at source ↗
Figure 3
Figure 3. Figure 3: A comparison between MLE and sparse Bayesian learn view at source ↗
Figure 4
Figure 4. Figure 4: An example of the application of the SSBL algorithm view at source ↗
Figure 5
Figure 5. Figure 5: Centres of the multilevel RBFs up to level 4. Each sy view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between Pure CFD, training DSMC, MMS-S view at source ↗
Figure 7
Figure 7. Figure 7: Comparison at the centrelines and upper wall betwe view at source ↗
Figure 8
Figure 8. Figure 8: Comparison between Pure CFD, training DSMC, MMS-S view at source ↗
Figure 9
Figure 9. Figure 9: Comparison at the centrelines and upper wall betwe view at source ↗
Figure 10
Figure 10. Figure 10: Comparison between Pure CFD, training DSMC, MMS- view at source ↗
Figure 11
Figure 11. Figure 11: Comparison at the centrelines and upper wall betw view at source ↗
Figure 12
Figure 12. Figure 12: Comparison between Pure CFD, training DSMC, MMS- view at source ↗
Figure 13
Figure 13. Figure 13: Comparison at the centrelines and upper wall betw view at source ↗
Figure 14
Figure 14. Figure 14: Comparison between Pure CFD, training DSMC, MMS- view at source ↗
Figure 15
Figure 15. Figure 15: Comparison at the centrelines and upper wall betw view at source ↗
Figure 16
Figure 16. Figure 16: Comparison between Pure CFD, training DSMC, MMS- view at source ↗
Figure 17
Figure 17. Figure 17: Comparison at the centrelines and upper wall betw view at source ↗
Figure 18
Figure 18. Figure 18: Comparison between Pure CFD, training DSMC, MMS- view at source ↗
Figure 19
Figure 19. Figure 19: Comparison at the centrelines and upper wall betw view at source ↗
read the original abstract

Hybrid methods for simulating rarefied gas flows reduce computational cost by coupling a particle-based model, typically the direct simulation Monte Carlo (DSMC) method, to a continuum-based solver, i.e. a computational fluid dynamics (CFD) code. However, widespread adoption of these methods is hindered by numerical instabilities caused by statistical noise and difficulties in applying them to complex, arbitrary geometries. To be effective, a hybrid framework must be robust to noise, reliable in not introducing errors to the flow physics, automated, and flexible enough for general spatial domains. Previous iterations of the micro-macro-surrogate-sparse (MMS-Sparse) method successfully addressed the first three requirements using Bayesian surrogate models to provide smooth, constitutive corrections to the CFD. However, they relied on global basis functions, limiting their applicability to relatively simple geometries. In this work, we address the fourth requirement - flexibility - by introducing a set of multilevel radial basis functions (RBFs) to represent the smooth corrections within the MMS-Sparse framework. Unlike global polynomials, multilevel RBFs can resolve broad and fine flow details locally, allowing the method to be applied to complex geometrical systems. We couple this approach with a finite-volume CFD solver (OpenFOAM) and validate it using the rarefied lid-driven cavity flow problem. This serves as a rigorous test case for spatially two-dimensional coupling. Our results demonstrate that this enhanced MMS-Sparse method produces estimates in good agreement with benchmarks while retaining the noise-robust and automated benefits of the Bayesian approach.

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 / 1 minor

Summary. The paper extends the MMS-Sparse Bayesian surrogate framework for hybrid DSMC-CFD coupling of rarefied flows by replacing global basis functions with multilevel radial basis functions (RBFs). This change is intended to enable application to complex arbitrary geometries while preserving noise robustness and automation. The method is coupled to OpenFOAM and validated on the 2D rarefied lid-driven cavity flow, with results claimed to agree with benchmarks.

Significance. If the multilevel RBF construction can be shown to deliver accurate constitutive corrections without introducing interpolation errors or coupling instabilities on non-rectangular domains, the work would meaningfully advance hybrid methods by removing a key geometric restriction from prior MMS-Sparse versions. The retention of the Bayesian noise-robustness property is a clear strength.

major comments (2)
  1. [§4 and Abstract] §4 (Numerical Results) and Abstract: The validation is performed exclusively on the lid-driven cavity in a rectangular domain. This geometry does not exercise the claimed flexibility for complex or arbitrary boundaries, leaving untested whether multilevel RBFs avoid the interpolation errors or instabilities that previously limited global bases.
  2. [Abstract and §4] Abstract and §4: No quantitative error metrics (e.g., L2 or L∞ norms versus benchmark solutions), convergence rates, or explicit measures of statistical-noise impact are supplied to substantiate the statements of “good agreement” and retained “noise robustness.”
minor comments (1)
  1. [§3] The description of the multilevel RBF construction (partitioning, overlap, and conditioning) would benefit from an explicit algorithmic outline or pseudocode to aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. We address each major comment point by point below, with proposed revisions to the manuscript where the comments identify areas for improvement.

read point-by-point responses

  1. Referee: [§4 and Abstract] §4 (Numerical Results) and Abstract: The validation is performed exclusively on the lid-driven cavity in a rectangular domain. This geometry does not exercise the claimed flexibility for complex or arbitrary boundaries, leaving untested whether multilevel RBFs avoid the interpolation errors or instabilities that previously limited global bases.

    Authors: We agree that the validation case is restricted to the rectangular lid-driven cavity. This geometry was selected as a standard, well-documented benchmark for 2D rarefied hybrid coupling that permits direct quantitative comparison with existing DSMC and hybrid results. The multilevel RBF construction is intended to remove the geometric restrictions of global bases by providing local support that can be adapted to arbitrary boundaries within the OpenFOAM finite-volume framework. Nevertheless, we acknowledge that the absence of a non-rectangular test case leaves the claims about interpolation errors and coupling stability on complex domains unverified in the present work. In the revised manuscript we will temper the language in the abstract and §4 to clarify the current validation scope and add a concise discussion of how the local RBF property enables extension to irregular domains, without overstating the demonstrated flexibility. revision: partial

  2. Referee: [Abstract and §4] Abstract and §4: No quantitative error metrics (e.g., L2 or L∞ norms versus benchmark solutions), convergence rates, or explicit measures of statistical-noise impact are supplied to substantiate the statements of “good agreement” and retained “noise robustness.”

    Authors: We thank the referee for highlighting this point. The current manuscript presents agreement primarily through visual overlay of profiles and contour plots. To strengthen the evidence, the revised §4 will include explicit quantitative metrics: L2 and L∞ norms of velocity and temperature fields relative to benchmark solutions, observed convergence behavior with respect to DSMC sample size, and a direct assessment of how statistical noise in the DSMC data propagates through the Bayesian surrogate. These additions will be supported by new tables or figures as appropriate. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper extends the prior MMS-Sparse Bayesian surrogate framework by replacing global basis functions with multilevel RBFs to enable complex-geometry coupling. No equations, predictions, or first-principles results are shown to reduce by construction to fitted inputs, self-definitions, or self-citation chains. The lid-driven cavity validation is presented as an independent benchmark test rather than a tautological output. All load-bearing steps (surrogate construction, noise robustness, OpenFOAM coupling) retain independent content and are not forced by renaming or ansatz smuggling.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract does not detail any free parameters, axioms, or invented entities; the approach builds on the existing MMS-Sparse framework using standard surrogate modeling techniques without introducing new postulated objects.

pith-pipeline@v0.9.0 · 5590 in / 1186 out tokens · 37164 ms · 2026-05-08T01:30:00.001707+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

78 extracted references · 3 canonical work pages

  1. [1]

    General physics: mechanics and molecular physics

    Lev Davidovich Landau, Aleksandr Il’ich Akhiezer, and E M Lifshitz. General physics: mechanics and molecular physics . Pergamon Press, 1967

    1967

  2. [2]

    Flow of rarefied gases

    Samuel A Schaaf and Paul L Chambré. Flow of rarefied gases . Vol. 8. Princeton University Press Princeton, 1961

    1961

  3. [3]

    Molecular gas dynamics and the direct simulation of gas flows

    Graeme A Bird. Molecular gas dynamics and the direct simulation of gas flows . Oxford university press, 1994

    1994

  4. [4]

    Statistical error in particle simulations of h ydrody- namic phenomena

    Nicolas G Hadjiconstantinou, Alejandro L Garcia, Marti n Z Bazant, and Gang He. “Statistical error in particle simulations of h ydrody- namic phenomena”. In: Journal of computational physics 187.1 (2003), pp. 274–297

    2003

  5. [5]

    Microflows and nanoflows: fundamentals and simulation

    George Karniadakis, Ali Beskok, and Narayan Aluru. Microflows and nanoflows: fundamentals and simulation . Vol. 29. Springer Science & Business Media, 2006. 40

    2006

  6. [6]

    Two-dimensional hybr id contin- uum/particle approach for rarefied flows

    Dean Wadsworth and Daniel Erwin. “Two-dimensional hybr id contin- uum/particle approach for rarefied flows”. In: 23rd Plasmadynamics and Lasers Conference . 1992, p. 2975

    1992

  7. [7]

    Molecular dynami cs– continuum hybrid computations: a tool for studying complex fluid flows

    Sean T O’connell and Peter A Thompson. “Molecular dynami cs– continuum hybrid computations: a tool for studying complex fluid flows”. In: Physical Review E 52.6 (1995), R5792

    1995

  8. [8]

    Adaptive mesh and algorithm refinement using direct simula- tion Monte Carlo

    Alejandro L Garcia, John B Bell, William Y Crutchfield, an d Berni J Alder. “Adaptive mesh and algorithm refinement using direct simula- tion Monte Carlo”. In: Journal of computational Physics 154.1 (1999), pp. 134–155

    1999

  9. [9]

    A hybrid part icle- continuum method applied to shock waves

    Thomas E Schwartzentruber and Iain D Boyd. “A hybrid part icle- continuum method applied to shock waves”. In: Journal of Compu- tational Physics 215.2 (2006), pp. 402–416

    2006

  10. [10]

    A hybrid continuum-pa rticle solver for unsteady rarefied gas flows

    Sarantis Pantazis and H Rusche. “A hybrid continuum-pa rticle solver for unsteady rarefied gas flows”. In: Vacuum 109 (2014), pp. 275–283

    2014

  11. [11]

    Coupling heterogeneous continuum-particl e fields to simulate non-isothermal microscale gas flows

    Stephanie Y Docherty, Matthew K Borg, Duncan A Lockerby , and Jason M Reese. “Coupling heterogeneous continuum-particl e fields to simulate non-isothermal microscale gas flows”. In: International Journal of Heat and Mass Transfer 98 (2016), pp. 712–727

    2016

  12. [12]

    Simulation of the head-disk interface gap using a hybrid multi-scale method

    Benzi John, Duncan A Lockerby, Alexander Patronis, and David R Emerson. “Simulation of the head-disk interface gap using a hybrid multi-scale method”. In: Microfluidics and Nanofluidics 22.9 (2018), p. 106

    2018

  13. [13]

    Particle-based hybrid and multiscale methods for nonequi librium gas flows

    Jun Zhang, Benzi John, Marcel Pfeiffer, Fei Fei, and Dong sheng Wen. “Particle-based hybrid and multiscale methods for nonequi librium gas flows”. In: Advances in Aerodynamics 1.1 (2019), p. 12

    2019

  14. [14]

    hybridDCFoam: A coupled DSMC/Navier–Stokes–Fourier solver for steady-state mult iscale rar- efied gas flows

    Nikolaos Vasileiadis and Craig White. “hybridDCFoam: A coupled DSMC/Navier–Stokes–Fourier solver for steady-state mult iscale rar- efied gas flows”. In: Advances in Engineering Software 193 (2024), p. 103669

    2024

  15. [15]

    Carte- sian grid-based hybrid NS-DSMC methodology for continuum- rarefied gas flows around complex geometries

    Vinay Kumar, UV Bhandarkar, RK Singh, and Atul Sharma. “ Carte- sian grid-based hybrid NS-DSMC methodology for continuum- rarefied gas flows around complex geometries”. In: Numerical Heat Transfer, Part B: Fundamentals 85.7 (2024), pp. 883–905. 41

    2024

  16. [16]

    Verification of a Modular DSMC-CFD Simula - tion Method Implementation for Expanding Rarefied Flows

    Ioannis Nompelis, Michael D Kroells, Thomas Schwartze ntruber, and Graham V Candler. “Verification of a Modular DSMC-CFD Simula - tion Method Implementation for Expanding Rarefied Flows”. I n: AIAA SCITECH 2025 Forum . 2025, p. 0951

    2025

  17. [17]

    A DSMC-CFD coupling method using surrogate modelling for low-speed ra refied gas flows

    Giorgos Tatsios, Arun K Chinnappan, Arshad Kamal, Niko s Vasileiadis, Stephanie Y Docherty, Craig White, Livio Gibelli, Matthew K Borg, James R Kermode, and Duncan A Lockerby. “A DSMC-CFD coupling method using surrogate modelling for low-speed ra refied gas flows”. In: Journal of Computational Physics 520 (2025), p. 113500

    2025

  18. [18]

    Bayesian regression-based continuum- particle method for low-speed rarefied flow: Application to unsteady P oiseuille flow

    Arun Kumar Chinnappan, Arshad Kamal, James R Kermode, a nd Duncan A Lockerby. “Bayesian regression-based continuum- particle method for low-speed rarefied flow: Application to unsteady P oiseuille flow”. In: Physics of Fluids 37.3 (2025)

    2025

  19. [19]

    Denoisi ng of direct simulation Monte Carlo data using proper orthogonal decompo- sition technique

    Rakesh Kumar, Arun K Chinnappan, and MV Akhil. “Denoisi ng of direct simulation Monte Carlo data using proper orthogonal decompo- sition technique”. In: Journal of Spacecraft and Rockets 55.4 (2018), pp. 841–847

    2018

  20. [20]

    Accel- erating multiscale modelling of fluids with on-the-fly Gauss ian process regression

    David Stephenson, James R Kermode, and Duncan A Lockerb y. “Accel- erating multiscale modelling of fluids with on-the-fly Gauss ian process regression”. In: Microfluidics and Nanofluidics 22.12 (2018), p. 139

    2018

  21. [21]

    Shape-informed surrogate models based on signed dis tance func- tion domain encoding

    Linying Zhang, Stefano Pagani, Jun Zhang, and Francesc o Regaz- zoni. “Shape-informed surrogate models based on signed dis tance func- tion domain encoding”. In: Journal of Computational Physics (2025), p. 114178

    2025

  22. [22]

    Data-Driven Surrog ate Modeling of DSMC Solutions Using Deep Neural Networks

    Ehsan Roohi and Ahmad Shoja-sani. “Data-Driven Surrog ate Modeling of DSMC Solutions Using Deep Neural Networks”. In: arXiv preprint arXiv:2506.22453 (2025)

    work page arXiv 2025

  23. [23]

    Accelerating Kinetic Fokker-Planck Simulations via a GPU-Native Deep Neural Network Surrogate: Application to R ar- efied Internal and Hypersonic External Flows

    Ehsan Roohi. Accelerating Kinetic Fokker-Planck Simulations via a GPU-Native Deep Neural Network Surrogate: Application to R ar- efied Internal and Hypersonic External Flows . Jan. 2026. doi: 10.2139/ssrn.6203942

    work page doi:10.2139/ssrn.6203942 2026

  24. [24]

    Pattern recognition and machine learning

    Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning . Vol. 4. 4. Springer, 2006. 42

    2006

  25. [25]

    The driven cavity flow over the whole range of the Knudsen number

    Stergios Naris and Dimitris Valougeorgis. “The driven cavity flow over the whole range of the Knudsen number”. In: Physics of fluids 17.9 (2005)

    2005

  26. [26]

    Investig ation of heat and mass transfer in a lid-driven cavity under nonequil ibrium flow conditions

    Benzi John, Xiao-Jun Gu, and David R Emerson. “Investig ation of heat and mass transfer in a lid-driven cavity under nonequil ibrium flow conditions”. In: Numerical Heat Transfer, Part B: Fundamentals 58.5 (2010), pp. 287–303

    2010

  27. [27]

    Sound propagation through a rarefied gas in rect angular chan- nels

    Lei Wu. “Sound propagation through a rarefied gas in rect angular chan- nels”. In: Physical Review E 94.5 (2016), p. 053110

    2016

  28. [28]

    Oscillatory rare fied gas flow inside a three dimensional rectangular cavity

    Peng Wang, Wei Su, and Yonghao Zhang. “Oscillatory rare fied gas flow inside a three dimensional rectangular cavity”. In: Physics of fluids 30.10 (2018)

    2018

  29. [29]

    Non- linear oscillatory rarefied gas flow inside a rectangular cav ity

    Peng Wang, Lianhua Zhu, Wei Su, Lei Wu, and Yonghao Zhang . “Non- linear oscillatory rarefied gas flow inside a rectangular cav ity”. In: Phys- ical Review E 97.4 (2018), p. 043103

    2018

  30. [30]

    Heat and mass transfer of oscillatory lid-driven cavity flow in the co ntinuum, transition and free molecular flow regimes

    Peng Wang, Wei Su, Lianhua Zhu, and Yonghao Zhang. “Heat and mass transfer of oscillatory lid-driven cavity flow in the co ntinuum, transition and free molecular flow regimes”. In: International Journal of Heat and Mass Transfer 131 (2019), pp. 291–300

    2019

  31. [31]

    Nu- merical modeling of the heat and mass transfer of rarefied gas flows in a double-sided oscillatory lid-driven cavity

    Qing He, Shi Tao, Liang Wang, Jiechao Chen, and Xiaoping Yang. “Nu- merical modeling of the heat and mass transfer of rarefied gas flows in a double-sided oscillatory lid-driven cavity”. In: International Journal of Heat and Mass Transfer 230 (2024), p. 125788

    2024

  32. [32]

    Varia nce reduction for Monte Carlo solutions of the Boltzmann equation

    Lowell L Baker and Nicolas G Hadjiconstantinou. “Varia nce reduction for Monte Carlo solutions of the Boltzmann equation”. In: Physics of Fluids 17.5 (2005)

    2005

  33. [33]

    A l ow- variance deviational simulation Monte Carlo for the Boltzm ann equa- tion

    Thomas MM Homolle and Nicolas G Hadjiconstantinou. “A l ow- variance deviational simulation Monte Carlo for the Boltzm ann equa- tion”. In: Journal of Computational Physics 226.2 (2007), pp. 2341– 2358

    2007

  34. [34]

    Low- variance direct Monte Carlo simulations using importance w eights

    Husain A Al-Mohssen and Nicolas G Hadjiconstantinou. “ Low- variance direct Monte Carlo simulations using importance w eights”. In: ESAIM: Modélisation mathématique et analyse numérique 44.5 (2010), pp. 1069–1083. 43

    2010

  35. [35]

    A varian ce-reduced di- rect Monte Carlo simulation method for solving the Boltzman n equa- tion over a wide range of rarefaction

    Mohsen Sadr and Nicolas G Hadjiconstantinou. “A varian ce-reduced di- rect Monte Carlo simulation method for solving the Boltzman n equa- tion over a wide range of rarefaction”. In: Journal of Computational Physics 472 (2023), p. 111677

    2023

  36. [36]

    Va riance re- duction for Fokker–Planck based particle Monte Carlo schem es

    M Hossein Gorji, Nemanja Andric, and Patrick Jenny. “Va riance re- duction for Fokker–Planck based particle Monte Carlo schem es”. In: Journal of Computational Physics 295 (2015), pp. 644–664

    2015

  37. [37]

    Varianc e reduced par- ticle solution of the Fokker-Planck equation with applicat ion to rarefied gas and plasma dynamics

    Mohsen Sadr and Nicolas G Hadjiconstantinou. “Varianc e reduced par- ticle solution of the Fokker-Planck equation with applicat ion to rarefied gas and plasma dynamics”. In: Journal of Computational Physics 492 (2023), p. 112402

    2023

  38. [38]

    A time-relaxed Monte Carlo method preserving the Navier- Stokes asymptotics

    Fei Fei. “A time-relaxed Monte Carlo method preserving the Navier- Stokes asymptotics”. In: Journal of Computational Physics 486 (2023), p. 112128

    2023

  39. [39]

    A multis cale stochastic particle method based on the Fokker-Planck mode l for nonequilibrium gas flows

    Ziqi Cui, Kaikai Feng, Qihan Ma, and Jun Zhang. “A multis cale stochastic particle method based on the Fokker-Planck mode l for nonequilibrium gas flows”. In: Journal of Computational Physics 520 (2025), p. 113458

    2025

  40. [40]

    Boosting the conve rgence of DSMC by GSIS

    Liyan Luo, Qi Li, Fei Fei, and Lei Wu. “Boosting the conve rgence of DSMC by GSIS”. In: Journal of Computational Physics 532 (2025), p. 113959

    2025

  41. [41]

    A Denois ing Multi- scale Particle Method for Nonequilibrium Flow Simulations

    Hao Yang, Kaikai Feng, Ziqi Cui, and Jun Zhang. “A Denois ing Multi- scale Particle Method for Nonequilibrium Flow Simulations ”. In: Jour- nal of Computational Physics (2025), p. 114096

    2025

  42. [42]

    Statistical simulation of low -speed rarefied gas flows

    Jing Fan and Ching Shen. “Statistical simulation of low -speed rarefied gas flows”. In: Journal of Computational Physics 167.2 (2001), pp. 393– 412

    2001

  43. [43]

    Fast marginal likel ihood max- imisation for sparse Bayesian models

    Michael E Tipping and Anita C Faul. “Fast marginal likel ihood max- imisation for sparse Bayesian models”. In: International workshop on artificial intelligence and statistics . PMLR. 2003, pp. 276–283

    2003

  44. [44]

    Direct simulation method based on BGK equation

    Jun Li. “Direct simulation method based on BGK equation ”. In: AIP Conference Proceedings. Vol. 1333. 1. American Institute of Physics. 2011, pp. 283–288. 44

    2011

  45. [45]

    A comparative study of the DSBGK and DVM methods for low-spe ed rarefied gas flows

    Minh Tuan Ho, Jun Li, Lei Wu, Jason M Reese, and Yonghao Zh ang. “A comparative study of the DSBGK and DVM methods for low-spe ed rarefied gas flows”. In: Computers & Fluids 181 (2019), pp. 143–159

    2019

  46. [46]

    A unified st ochastic particle Bhatnagar-Gross-Krook method for multiscale gas flows

    Fei Fei, Jun Zhang, Jing Li, and ZhaoHui Liu. “A unified st ochastic particle Bhatnagar-Gross-Krook method for multiscale gas flows”. In: Journal of Computational Physics 400 (2020), p. 108972

    2020

  47. [47]

    A Navier-Stokes asymptotic preserving direc t simulation Monte Carlo method for multi-species gas flows

    Fei Fei. “A Navier-Stokes asymptotic preserving direc t simulation Monte Carlo method for multi-species gas flows”. In: Journal of Com- putational Physics (2025), p. 114196

    2025

  48. [48]

    Monte-Carlo simulation in an engineering con text

    GA Bird. “Monte-Carlo simulation in an engineering con text”. In: Progress in Astronautics and Aeronautics 74 (1981), pp. 239–255

    1981

  49. [49]

    Direct simulation Monte Carlo o n petaflop supercomputers and beyond

    Steve J Plimpton, SG Moore, A Borner, AK Stagg, TP Koehle r, JR Torczynski, and MA Gallis. “Direct simulation Monte Carlo o n petaflop supercomputers and beyond”. In: Physics of Fluids 31.8 (2019)

    2019

  50. [50]

    https://sparta.github.io

    SPARTA Direct Simulation Monte Carlo (DSMC) Simulator . https://sparta.github.io. Accessed: 2025-05-31

    2025

  51. [51]

    A tensorial approach to computational continuum mechanic s using object-oriented techniques

    Henry G Weller, Gavin Tabor, Hrvoje Jasak, and Christer Fureby. “A tensorial approach to computational continuum mechanic s using object-oriented techniques”. In: Computers in physics 12.6 (1998), pp. 620–631

    1998

  52. [52]

    https://doc.cfd.direct/openfoam/user-guide-v9

    OpenFOAM v9 User Guide. https://doc.cfd.direct/openfoam/user-guide-v9. Accessed: 2025-05-31

    2025

  53. [53]

    Computational methods for fluid dynamics

    Joel H Ferziger and Milovan Perić. Computational methods for fluid dynamics. Springer, 2002

    2002

  54. [54]

    Sparse Bayesian learning and the re levance vec- tor machine

    Michael E Tipping. “Sparse Bayesian learning and the re levance vec- tor machine”. In: Journal of machine learning research 1.Jun (2001), pp. 211–244

    2001

  55. [55]

    Analysis of sparse Bay esian learning

    Anita Faul and Michael Tipping. “Analysis of sparse Bay esian learning”. In: Advances in neural information processing systems 14 (2001)

    2001

  56. [56]

    Data and Codes for the pa- per ’Bayesian regression-based continuum-particle metho d for low-speed rarefied flow: application to unsteady Poiseuille flow’

    Arun Kumar Chinnappan, Arshad Kamal, James R Ker- mode, and Duncan A Lockerby. Data and Codes for the pa- per ’Bayesian regression-based continuum-particle metho d for low-speed rarefied flow: application to unsteady Poiseuille flow’ . https://doi.org/10.5281/zenodo.14968283. 2025. 45

    work page doi:10.5281/zenodo.14968283 2025

  57. [57]

    Meshfree approximation methods with Matlab (With Cd-rom)

    Gregory E Fasshauer. Meshfree approximation methods with Matlab (With Cd-rom) . Vol. 6. World Scientific Publishing Company, 2007

    2007

  58. [58]

    Learning from data

    Yaser S Abu-Mostafa, Malik Magdon-Ismail, and Hsuan-T ien Lin. Learning from data . Vol. 4. AMLBook New York, 2012

    2012

  59. [59]

    Six strategies for defeating the Runge Phe nomenon in Gaussian radial basis functions on a finite interval

    John P Boyd. “Six strategies for defeating the Runge Phe nomenon in Gaussian radial basis functions on a finite interval”. In: Computers & Mathematics with Applications 60.12 (2010), pp. 3108–3122

    2010

  60. [60]

    Determining of Flow Non - equilibrium Regions using the Kolmogorov-Smirnov Criteri on

    EV Titov, R Kumar, and DA Levin. “Determining of Flow Non - equilibrium Regions using the Kolmogorov-Smirnov Criteri on”. In: AIP Conference Proceedings. Vol. 1333. 1. American Institute of Physics. 2011, pp. 563–568

    2011

  61. [61]

    Monte Carlo sim ulation of the Taylor–Couette flow of a rarefied gas

    Stefan Stefanov and Carlo Cercignani. “Monte Carlo sim ulation of the Taylor–Couette flow of a rarefied gas”. In: Journal of Fluid Mechanics 256 (1993), pp. 199–213

    1993

  62. [62]

    A m ulti- scale method for micro/nano flows of high aspect ratio

    Matthew K Borg, Duncan A Lockerby, and Jason M Reese. “A m ulti- scale method for micro/nano flows of high aspect ratio”. In: Journal of Computational Physics 233 (2013), pp. 400–413

    2013

  63. [63]

    Bayesian inference in statistical analysis (Vol

    GE Box and George C Tiao. Bayesian inference in statistical analysis (Vol. 40). 2011

    2011

  64. [64]

    Bayesian interpolation

    David JC MacKay. “Bayesian interpolation”. In: Neural computation 4.3 (1992), pp. 415–447

    1992

  65. [65]

    Bayesian data analysis

    Andrew Gelman, John B Carlin, Hal S Stern, and Donald B Ru bin. Bayesian data analysis . Chapman and Hall/CRC, 1995. Appendix A. The Pressure-Implicit with Splitting of Operat ors al- gorithm For the macro-model, we use the icoFoam solver to solve the in - compressible, transient N-S equations (2) - (3), with Diric hlet boundary conditions for the flow vel...

    1995

  66. [66]

    Set initial conditions for flow velocity and scalar pressu re

  67. [67]

    Start procedure to compute next time step values

  68. [68]

    Execute the PISO loop until updates in the corrected solut ion are suf- ficiently small: (a) Solve the momentum equation (2) for the flow velocity foun d by finite-volume discretisation ( Predictor Step ). (b) Given flow velocity and fluxes at the faces of the cells, sol ve the pressure-correction equation, found by discretising the c ontinuity equation (3), f...

  69. [69]

    Store converged flow velocity and scalar pressure values, as well as associated fluxes, at current time-step

  70. [70]

    Otherwise, go to St ep 2 for the next time-step

    Terminate here if final time is reached. Otherwise, go to St ep 2 for the next time-step. Appendix B. Gamma hyperpriors in sparse Bayesian learning When the noise of the DSMC is very high, we found that the SSBL algorithm can produce over-fitted estimates that are unsuit able for use in a CFD model. Here, we briefly detail a strategy to overcome the o ver-fit...

  71. [71]

    Set threshold wthres for the weights wi based on known parameters (e.g., target vector ˆd and design matrix Ψ ): wthres = 10 − 4 × ||ˆd||/ ||Ψ ||

  72. [72]

    Set also thresholds for convergence in α and β (i.e., 10− 4)

  73. [73]

    Initialise α and β from gamma distributions of the form Γ(2 , 2/A i) and Γ(2 , 2/A ), where Ai and A are some large parameters. 47

  74. [74]

    Calculate mean m and covariance Σ as before, using equations (15) and (16)

  75. [75]

    Update α and β , taking into account the gamma distribution parame- ters a1 = a2 = 2 and b1 = 2/A i, b 2 = 2/A : α new i = γi + 2a1 m2 i + 2b1 , (B.1) (β new)− 1 = ||ˆd − Ψm ||2 + 2b2 NxNy − ∑ i γi + 2a2 , (B.2) where we have defined γi = 1 − α iΣ ii as in [64]

  76. [76]

    Prune components of α whose corresponding mean |mi|< w thres

  77. [77]

    In this algorithm, in Steps 3 and 5, we consider A = 5 × 102 for the inverse-variance β and Ai = 5 × 102 × i2 for level i corresponding to α i

    Repeat Step 4 to Step 6 until convergence criteria is met fo r α and β . In this algorithm, in Steps 3 and 5, we consider A = 5 × 102 for the inverse-variance β and Ai = 5 × 102 × i2 for level i corresponding to α i. With these choices, the resulting posterior distributions are a ssumed to depend on the mode of the dataset [65]. Appendix B.1. Low-noise vs...

  78. [78]

    05 & M = 0

    for high-noise DSMC data at Kn = 0. 05 & M = 0. 1. 50 Figure C.22: Regression estimates using SSBL, with β − 1 updates, up to and including level 4 for Benchmark DSMC data at Kn = 0. 05 & M = 0. 1. 51 with u, v & τxy as the target variables. Figure C.22 shows that considering RBFs up to level 4 (i.e., 17 × 17 = 289 RBFs) is sufficient to capture the require...