Data-Driven Surrogate Modeling of DSMC Solutions Using Deep Neural Networks

Ahmad Shoja-Sani; Ehsan Roohi

REVIEW 4 major objections 5 minor 23 references

Data-Driven Surrogate Modeling of DSMC Solutions Using Deep Neural Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that deep neural networks can imitate Direct Simulation Monte Carlo solutions for rarefied-gas flows, turning minute-to-hour DSMC runs into millisecond predictions across equilibrium distributions, shock profiles, and…

desk verdict The paper's headline extrapolation claim is contradicted by its own dataset description; the underlying engineering is real but the quantitative claims are unverified. read the letter →

arxiv 2506.22453 v1 pith:CZJCZ3HG submitted 2025-06-14 physics.comp-ph physics.flu-dyn

Ehsan Roohi , Ahmad Shoja-sani This is my paper · ORCID

classification physics.comp-phphysics.flu-dyn PACS 47.45.-n

keywords DeepneuralnetworksurrogateDirectSimulationMonteCarloRarefiedgasdynamicsMaxwell-BoltzmanndistributionFourierfeaturemappingKnudsennumberShockwaveextrapolationLid-drivencavityflow

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that deep neural networks can serve as fast, accurate surrogates for Direct Simulation Monte Carlo (DSMC) in rarefied-gas flows. Three test problems are used: reproducing argon's Maxwell-Boltzmann speed distribution over temperatures 200-650 K; predicting one-dimensional shock-wave profiles for Mach numbers beyond those seen in training; and reconstructing two-dimensional velocity and temperature fields in a lid-driven cavity at untrained Knudsen numbers. The reported payoffs are large: inference times of milliseconds instead of tens of minutes, and spatial errors under 2% for the cavity fields. If these results hold, surrogate models of this kind would make parametric studies, uncertainty quantification, and design optimization practical for flows in which the continuum Navier-Stokes equations break down.

What carries the argument

The machinery is threefold, each part tied to one failure mode. First, explicit physical constraints in preprocessing: appending the zero-speed/zero-density point to the Maxwell-Boltzmann training set, and rescaling variables so all targets are non-negative, enforces known limits a naive network would otherwise violate. Second, Fourier feature mapping: the spatial coordinate is lifted to a high-dimensional set of sine and cosine features with a (sometimes learnable) frequency matrix, which lets a network represent steep shock gradients as combinations of smooth basis functions rather than piecewise-linear ReLU steps; tanh or swish activations, dropout, and L2 regularization keep the extrapolation stable. Third, the family-of-experts scheme: separate specialist networks are trained at discrete Knudsen numbers, and predictions at untrained Kn are formed by weights derived from the bracketing specialists. The weight formula $w = \ln(Kn_{\mathrm{test}}/Kn_1)/\ln(Kn_2/Kn_1)$ combines the two specialists' outputs in log-Knudsen space.

What would settle it

Re-train the optimized shock architecture on a set of files that provably excludes Mach 2.0, then compute the pointwise error of density, velocity, and temperature at Mach 2.0 and 2.5 against the Bird DSMC1S reference; if Mach 2.0 had been in the training file, the agreement reported for M=2.0 would be interpolation, not extrapolation, and the M=2.5 result alone would carry the extrapolation claim.

Watch

Extended reading notes

Core claim

The central discovery the authors report is that a simple feed-forward neural network, given the right input encoding and physical constraints, can reproduce the output of DSMC closely enough to stand in for it. For the equilibrium test, appending the physical boundary point ($v=0$, $P=0$) to the training data enforces the correct low-speed limit and brings the mean-squared error below $10^{-5}$; an additional non-dimensionalization based on the thermal speed turns temperature extrapolation into interpolation on a single universal curve. For shocks, a Fourier-feature encoding of position plus a sigmoid-bounded output lets a network trained at Mach 1.4-1.9 predict Mach 2 and 2.5 profiles that the paper says agree with DSMC in position, amplitude, and slope. For the lid-driven cavity, a family of experts, one specialist network per Knudsen number, combined with interpolation of their outputs in log-Knudsen space recovers the full 2-D fields at unseen Kn with less than 2% spatial error.

Load-bearing premise

The paper's extrapolation claim rests on the unstated assumption that the optimized shock network was trained only at Mach 1.4-1.9 and that the Mach 2.0 case listed in Section 2.2.1 as a training file was excluded from that model's training set.

Editorial extensions

If this is right

Parametric sweeps over temperature, Mach number, or Knudsen number can be evaluated in milliseconds once the surrogate is trained, making many-query design studies practical that would be far slower with DSMC alone.
The boundary-point and non-dimensionalization tricks are transferable preprocessing steps; the relaxation test shows they can convert temperature extrapolation into a simple curve-fitting task.
The family-of-experts plus log-interpolation recipe can cover a wide Knudsen range with modest data, since each specialist only needs to learn one spatial mapping per regime.
The paper states the framework is intended as a template for non-equilibrium phenomena, gas mixtures, and design optimization workflows, and expects it to accelerate uncertainty quantification and multi-scale micro/nano-flow modeling.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension the paper does not carry out is a held-out Mach sweep beyond 2.5; if the same architecture extrapolates to Mach 3 or 4, the claim of robust out-of-training generalization would be much stronger.
A broader reading is that log-space regime interpolation could transfer to other dimensionless parameters, such as Reynolds number or pressure ratio, in other particle or lattice-Boltzmann settings.
A natural follow-up, suggested by the boundary-point trick, would be to hard-constrain the first few moments of the predicted distribution, which could reduce the need for specialist models at fine Knudsen spacing.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper applies standard DNN-surrogate techniques to DSMC data, and the most useful part is the cavity 'family-of-experts' idea: train per-Kn networks and interpolate in log-space. That's a sensible engineering trick for a problem where a single network struggles. The DSMC data generation work looks careful, and the LightGBM comparison is a nice sanity check. So there is genuine craft here.

But the central generalization claim does not survive contact with the paper's own Section 2.2.1. The abstract says the shock network was trained on Mach 1.4–1.9 and extrapolates to Mach 2.0 and 2.5. Section 2.2.1, however, explicitly lists Mach 2.0 among the training Mach numbers, with 1.7 held out as the interpolation test. The later section on the "optimized" model never states that the Mach 2.0 file was discarded. If it was not discarded, the M=2.0 result is interpolation, not extrapolation, and the "robust out-of-training generalization" claim collapses. That is the load-bearing problem.

Other soft spots: the <2% spatial error for the cavity is asserted in the abstract but no error metric is computed anywhere; the shock results are shown as figures with no quantified error. The dimensionless Maxwell-Boltzmann transformation is mis-stated: defining s = v_char/v where v_char = sqrt(m k_B T) gives a reciprocal speed, and the displayed P(s) does not integrate to 1. The log-interpolation equation is missing—Eq. (2) is empty. And the "PINN" for the equilibrium distribution is just a coordinate rescaling, not a physics-informed neural network; calling it a PINN overstates it.

None of these are unfixable. The paper is coherent in its overall approach, and the cavity specialist methodology is worth preserving. But as written, the abstract oversells and the dataset contradiction undermines the main result. A serious referee could push the authors to state exactly which Mach numbers were in the final training set, to report real errors, and to correct the PDF formula. I'd send it to peer review with the expectation of major revision, but I would not desk-reject it—there is a usable idea underneath.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes three deep neural network surrogates for DSMC-based rarefied-gas simulations: (i) a fully connected network that learns the argon Maxwell–Boltzmann speed distribution as a function of speed and temperature, with an added physical boundary point at v=0 and a 'PINN' variant based on a claimed dimensionless scaling law; (ii) a Fourier-feature network for 1-D shock-wave profiles that is claimed to extrapolate from training Mach numbers 1.4–1.9 to Mach 2.0 and 2.5; and (iii) a 'family-of-experts' strategy for lid-driven cavity flow, in which specialist models at discrete Knudsen numbers are fused by logarithmic interpolation to predict intermediate Kn cases. The abstract claims near-perfect shock extrapolation and less than 2% spatial error for the cavity surrogate, with millisecond inference times.

Significance. If fully substantiated, the framework would be a useful contribution to many-query rarefied-gas design studies, and the paper contains some genuinely constructive elements: the explicit injection of the zero-speed boundary point, the modular specialist-model decomposition, and the attempt to test extrapolation beyond the training range are all valuable ideas. However, the quantitative headline claims are not supported by the evidence as printed: no error metrics are given for the shock extrapolation or cavity results, the dataset description contradicts the 'trained only up to Mach 1.9' claim, and the dimensionless Maxwell–Boltzmann formula in the physics-informed section is dimensionally and normalization-wise incorrect. These issues directly affect the central claims and require substantive revision.

major comments (4)

[Sec. 2.2.1 and Sec. 2.2.5.3] The extrapolation claim is not supported by the dataset description. Section 2.2.1 states that the model was trained on Mach numbers 1.4, 1.5, 1.6, 1.8, 1.9, and 2.0, with only Mach 1.7 held out; Sections 2.2.5.3–2.2.5.5 describe the optimized model as trained only up to Mach 1.9 but never state that the Mach 2.0 file was removed. Consequently, the M=2.0 'extrapolation' may actually be an interpolation result. The authors must state explicitly which Mach numbers were used in the optimized model's training set and, if M=2.0 was excluded, verify or retrain under that exclusion; if it was not excluded, the abstract's 'trained only on Mach numbers 1.4–1.9' is incorrect. In addition, no quantitative error is reported for the M=2.0 or M=2.5 predictions; the paper should report normalized errors (e.g., L2 or maximum error over the profile) for density, velocity, and temperature instead of relying only on visual agreement.
[Sec. 2.1.4 (second subsection of that number)] The dimensionless Maxwell–Boltzmann formula is incorrect. The paper defines s=vchar/v and states P(s)=2/π × s × exp(−s/2). This function is not a probability density: its integral over s from 0 to infinity is 8/π, not 1. Moreover, with the stated change of variables P(s)ds=P(v)dv and vchar=√(kT/m), direct substitution of v=vchar/s yields P(s)=√(2/π) s^{-4} exp(−1/(2s^2)), not the stated expression; equivalently, the conventional dimensionless speed uses s=v/vchar and gives P(s)=4/√π s^2 exp(−s^2). Because this formula is the stated basis of the 'physics-informed' scaling law and the universal-curve claim, the derivation must be corrected and the network target re-derived.
[Sec. 2.3.4 and Abstract] The headline claim of 'less than 2% spatial error' for the cavity surrogate is not evidenced in the manuscript. Section 2.3.4 provides only qualitative contour comparisons and line plots for Kn=0.05 and 0.5, with no definition of the spatial error norm, no numerical error values, and no error field. The interpolation formula referenced as Eq. (2) is also missing from the text. The authors should include the exact log-interpolation formula, define the spatial error metric, and report numerical errors or error maps for the test Knudsen numbers.
[Secs. 2.2.5.1, 2.2.5.5, and 2.3.3.1] The description of the Fourier-feature layer is internally inconsistent. Section 2.2.5.1 states that the mapping matrix B is randomly initialized from a normal distribution, while Section 2.2.5.5 refers to a learnable Fourier-feature mapping with scale 5.0, and Section 2.3.3.1 explicitly keeps B non-trainable. Since the learnability and scale of B are central to reproducing the shock extrapolation results, the manuscript must state unambiguously whether B is trainable in each model and how the scale hyperparameter is set.

minor comments (5)

[Throughout] There are duplicated section and figure numbers: Section 2.1.4 appears twice, and Figure 10 is used both for the architecture in Section 2.2.5.2 and for the M=2.0 results in Section 2.2.5.3. These need to be renumbered.
[References] Reference [13] (Wu et al.) is cited in the introduction but is missing from the reference list, and reference [16] contains garbled URL fragments; both need correction.
[Sec. 2.2.1] The data file is described as containing six columns including Position, Density, Velocity, Temperature, and Translational Temperature, but the network output is described as four quantities; the correspondence between the file columns and the network outputs should be clarified.
[Sec. 2.3.3.4] Equation (2), the logarithmic interpolation weight formula, is not actually printed in the manuscript even though the text refers to it; the equation should be added.
[Sec. 2.1.4.2] There are typographical errors and informal phrasing throughout (e.g., 'On the other hands'), and the inference-time speed-up is described only qualitatively; a copyedit and a concrete timing table would improve the presentation.

Circularity Check

1 steps flagged · score 2.0 of 10

No equation-level circularity; the Maxwell-Boltzmann 'extrapolation' is a coordinate rescaling of the known analytical law, and the shock M=2.0 'extrapolation' claim conflicts with the dataset description.

renaming known result [Sec. 2.1.4 (PINN for Maxwell-Boltzmann), universal-curve paragraph and extrapolation discussion.]
"Here, we suggest defining a characteristic speed, vchar=mkBT ... The crucial insight is that the resulting dimensionless distribution, P(s)=2/π×s×exp(−s/2), is a universal function that is independent of temperature. ... The challenging task of extrapolation in the temperature domain was not handled by the neural network, but rather by the analytical scaling laws applied during the pre- and post-processing steps."

The claimed out-of-training prediction at 900 K is produced by the analytical Maxwell-Boltzmann scaling law rather than by the network: the 'universal' dimensionless curve is Equation (1) after a change of variables, and the inverse transform reinserts the temperature dependence of the target distribution. Therefore the extrapolation result is equivalent, by construction, to evaluating the known physical formula at the new temperature; the DNN only fits the collapsed curve. The paper is explicit about this, so the issue is a mild renaming/over-claim rather than a hidden circular derivation.

full rationale

The core DNN-to-DSMC mappings for shocks and cavity flows are not circular: the networks are trained on DSMC profiles and evaluated at unseen Mach or Knudsen numbers, with no fitted parameter that is then renamed as a prediction; the cavity 'family-of-experts' result is an explicitly labeled log-space interpolation between specialist models, not an equation-level tautology. Self-citations [3,6-8,19-21] are background DSMC references and are not load-bearing for the neural-network claims. Two issues prevent a score of 0. First, the Maxwell-Boltzmann 'PINN' extrapolation to 900 K reduces to the analytical MB scaling law in rescaled coordinates, as the paper itself states; calling this a network extrapolation is a renaming of a known result. Second, Sec. 2.2.1 lists Mach 2.0 among training Mach numbers, while Sec. 2.2.5.3 and the Abstract say the model was 'trained only up to Mach 1.9'; the paper never states that the M=2.0 file was excluded from the optimized model, so the 'Mach 2.0 extrapolation' may be in-sample. This is a data-composition/validation inconsistency rather than an equation-level circular reduction, but it directly affects the headline generalization claim. Sec. 2.2.5.4 also describes iterative tuning against the M=2.5 case, which is a test-set selection concern, not an equation-level circularity. No Lean/Coq-style machine-checked support is invoked, but none is claimed.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The central claims rest on four domain assumptions about DSMC data quality, smoothness in Mach/Kn, and the validity of log interpolation; the PINN scaling claim is weakened by an incorrect formula.

free parameters (2)

Fourier feature scale (shock final model) = 5.0
Set by hand; paper says final model performance is sensitive to this scale (Sec. 2.2.5.5); no ablation.
Fourier feature scale (cavity specialists) = 2.5
Tuned to resolve spatial details without high-frequency noise (Sec. 2.3.3.1).

assumptions (6)

standard math Argon equilibrium speed distribution follows the Maxwell-Boltzmann formula.
Sec. 2.1.2 Eq. (1); standard statistical mechanics, not in dispute.
domain assumption DSMC reference data are statistically converged and effectively noise-free for training labels.
No convergence or sampling variance analysis is reported for any of the three datasets; the DSMC results are used as ground truth.
ad hoc to paper The dimensionless speed transform P(s)ds=P(v)dv collapses all equilibrium temperatures to one curve.
The paper states this transform (Sec. 2.1.4) but gives a non-normalized, dimensionally inconsistent formula, so the collapse is not established as written.
domain assumption Shock profiles depend smoothly and continuously on Mach number, so interpolation/extrapolation in M is well-posed.
Used implicitly in Sec. 2.2.5; no regularity or error-bound argument is provided.
ad hoc to paper Log-space interpolation of specialist network outputs is valid for intermediate Knudsen numbers.
The interpolation weight formula (Eq. 2) is missing and no validation with a held-out multi-Kn error metric is reported (Sec. 2.3.3.4).
domain assumption Fourier feature mappings with fixed or learnable frequency matrices can represent shock gradients at the selected scales.
Motivated by prior work [15-18], but the scale choice is empirical and not justified a priori (Sec. 2.2.5.1, 2.3.3.1).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Data-Driven Surrogate Modeling of DSMC Solutions Using Deep Neural Networks." pith.science (2026). https://pith.science/paper/CZJCZ3HG

@misc{pith2026250622453,
  author       = {Pith},
  title        = {Pith review of: Data-Driven Surrogate Modeling of DSMC Solutions Using Deep Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZJCZ3HG}},
  note         = {Machine review of arXiv:2506.22453}
}

read the original abstract

This study presents a deep neural network (DNN) framework that accelerates Direct Simulation Monte Carlo (DSMC) computations for rarefied-gas flows, while maintaining high physical fidelity. First, a fully connected deep neural network is trained on high-quality DSMC data for seven temperatures (200-650 K) to reproduce the Maxwell-Boltzmann speed distribution of argon. Injecting the physical boundary point into the training set enforces the correct low-speed limit. It reduces the mean-squared error to below 10^-5, thereby decreasing inference time from tens of minutes per DSMC run to milliseconds. For one-dimensional shock waves, a multi-output network equipped with learnable Fourier features learns the complete profiles of density, velocity, and temperature. Trained only on Mach numbers 1.4-1.9, it predicts a Mach 2 and 2.5 case with near-perfect agreement to DSMC, demonstrating robust out-of-training generalization. In a lid-driven cavity, the large parametric spread in Knudsen number is handled by a "family-of-experts" strategy: separate specialist models are trained at discrete Knudsen (Kn) values, and log-space interpolation fuses their outputs. This hybrid surrogate recovers the full 2-D velocity and temperature fields at unseen Kn with less than 2% spatial error. Key innovations include (i) explicit injection of physical constraints during data preprocessing, (ii) learnable Fourier feature mapping to capture steep shock gradients, and (iii) a modular expert-interpolation scheme to cover wide Knudsen ranges. Together, they establish a general recipe for trustworthy, rapid surrogate models that can be extended to non-equilibrium phenomena, gas mixtures, and design optimization workflows

Figures

Figures reproduced from arXiv: 2506.22453 by the authors.

**Figure 1.** The structure diagram of the employed neural network. The training process is as follows: The Adam optimizer was used with an initial learning rate of 0.001. Mean Squared Error (MSE) was used to quantify the difference between the model's predictions and the true values. Two key callbacks were employed to manage the training: ReduceLROnPlateau automatically reduces the learning rate if the validation loss stops impr… view at source ↗

**Figure 2.** The training and validation loss curves for the DNN model over 250 epochs [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 5.** VDF from LightGBM model for an unlearned temperature of 325.0 K compared to the analytical solution. As can be seen, the LightGBM model successfully learned the general unimodal, bell-curve shape of the Maxwell-Boltzmann distribution. However, upon closer inspection, clear discrepancies with the theoretical curve are evident. The prediction (green dashed line) exhibits two primary inaccuracies: First, Peak Amplitude… view at source ↗

Figures from the paper (10 more)

**Figure 6.** Figure 6: illustrates the training history of the employed DNN, showing the evolution of training and validation loss over 175 epochs. Both losses, measured in mean squared error (MSE), decrease rapidly during the initial epochs, indicating effective learning. From approximately…

**Figure 8.** Figure 8: The structure diagram of the employed neural network for [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

**Figure 9.** Figure 9: DNN prediction of the density, velocity, and temperature for the shock wave at an unlearned Mach number of 1.7. 2.2.5 An Optimized Neural Network for Shock Wave Extrapolation Here, we aim to develop a DNN capable of accurately extrapolating the velocity profile of the …

**Figure 10.** Figure 10: The structure diagram of the employed neural network based on [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

**Figure 10.** Figure 10: DNN prediction of the density, velocity, and temperature for the shock wave at an [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]

**Figure 11.** Figure 11: DNN prediction of the density, velocity, and temperature for the shock wave at an [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

**Figure 12.** Figure 12: The schematic of the cavity considered for DSMC simulations and DNN predictions. 2.3.2 Modeling Strategy: A Hybrid Surrogate Model for Parametric DSMC Simulations of Cavity Flow This section outlines a successful methodology for creating a high-fidelity surrogate mode…

**Figure 13.** Figure 13: The structure diagram of the employed neural network for the cavity problem. [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]

**Figure 14.** Figure 14: DSMC and DNN prediction for the cavity flow velocity and temperature, Kn=0.05 [PITH_FULL_IMAGE:figures/full_fig_p025_14.png]

**Figure 16.** Figure 16: DSMC and DNN contour field predictions for the cavity flow at Kn=0.5. [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

[1]

A., Molecular Gas Dynamics and the Direct Simulation of Gas Flows, Oxford Univ

Bird, G. A., Molecular Gas Dynamics and the Direct Simulation of Gas Flows, Oxford Univ. Press, 1994

work page 1994
[2]

A., The DSMC Method, 2013

Bird, G. A., The DSMC Method, 2013

work page 2013
[3]

Akhlaghi, H., Roohi, E., & Stefanov, S. (2023). A comprehensive review on micro-and nano- scale gas flow effects: Slip -jump phenomena, Knudsen paradox, thermally -driven flows, and Knudsen pumps, Physics Reports, 997, 1-60

work page 2023
[4]

Variable Soft Sphere molecular model for accurate rarefied gas flow simulation,

Koura, K. & Matsumoto, H., “Variable Soft Sphere molecular model for accurate rarefied gas flow simulation,” Phys. Fluids A, 3 (10), 2459–2465 (1991)

work page 1991
[5]

Collision sampling schemes for the Boltzmann equation and DSMC,

Ivanov, M. S. & Markelov, G. N., “Collision sampling schemes for the Boltzmann equation and DSMC,” USSR Comput. Math. Math. Phys., 24 (2), 84–96 (1985)

work page 1985
[6]

Goshayeshi, E

B. Goshayeshi, E. Roohi, S. Stefanov, DSMC simulation of hypersonic flows using an improved SBT-TAS technique, J. Comput. Phys. 303 (2015) 28–44

work page 2015
[7]

Shoja-sani, E

A. Shoja-sani, E. Roohi, S. Stefanov, Evaluation of the generalized Bernoulli trial-transient adaptive subcell ( GBT-TAS ) collision scheme in treating rarefied gas flows, Comput. Fluids. 213 (2020) 104740

work page 2020
[8]

Javani, E

M. Javani, E. Roohi, S. Stefanov, Symmetrized generalized and simplified Bernoulli-trials collision schemes in DSMC, Comput. Fluids. 272 (2024) 106188

work page 2024

Show all 23 references

[9]

Neural -network acceleration of the Boltzmann collision integral,

Xiao, H. & Frank, J. “Neural -network acceleration of the Boltzmann collision integral,” J. Comput. Phys. 407, 109246 (2020)

2020
[10]

A neural -network-based collision operator for DSMC,

Roberts, J., Mansour, N. N. & Laurence, D. “A neural -network-based collision operator for DSMC,” J. Comput. Phys. 470, 111541 (2022)

2022
[11]

Hybrid lattice -Boltzmann/NN-BGK solver for rarefied gas dynamics,

Zhang, L., Xu, H. & Chen, S. “Hybrid lattice -Boltzmann/NN-BGK solver for rarefied gas dynamics,” Phys. Fluids 35, 054105 (2023)

2023
[12]

Physics -informed neural networks for the spatially homogeneous Boltzmann equation,

Thapa, B. & Karniadakis, G. E. “Physics -informed neural networks for the spatially homogeneous Boltzmann equation,” Phys. Rev. Fluids 6, 124703 (2021). 30

2021
[14]

Deep-learning surrogate coupling for DSMC–CFD microflow simulations,

Tatsios, G., Markelov, G. & Ivanov, M. “Deep-learning surrogate coupling for DSMC–CFD microflow simulations,” J. Comput. Phys. 520, 113500 (2025)

2025
[15]

Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains,

Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng (2020), “Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains,” NeurIPS 2020

2020
[16]

A Deep Domain Decomposition Method Based on Fourier Features (F -D3M),

Sen Li, Yingzhi Xia, Yu Liu, and Qifeng Liao (2022), “A Deep Domain Decomposition Method Based on Fourier Features (F -D3M),” Computer Physics Communications bytez.com+14arxiv.org+14sciencedirect.com+14

2022
[17]

Deep Ritz Method with Fourier Feature Mapping: A Deep Learning Approach for Solving Variational Models of Microstructure,

Ensela Mema, Ting Wang, and Jaroslaw Knap (2025), “Deep Ritz Method with Fourier Feature Mapping: A Deep Learning Approach for Solving Variational Models of Microstructure,” arXiv preprint arxiv.org+11arxiv.org+11arxiv.org+11

2025
[18]

SASNet: Spatially-Adaptive Sinusoidal Neural Networks,

Haoan Feng, Diana Aldana, Tiago Novello, and Leila De Floriani (2025), “SASNet: Spatially-Adaptive Sinusoidal Neural Networks,” arXiv preprint aimodels.fyi+9arxiv.org+9paperswithcode.com+9

2025
[19]

Mohammadzadeh, A., Roohi, E., Niazmand, H., Stefanov, S., & Myong, R. S. (2012). Thermal and second-law analysis of a micro -or nanocavity using direct-simulation Monte Carlo. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 85(5), 056310

2012
[20]

Akhlaghi, H., Roohi, E., & Stefanov, S. (2018). Ballistic and collisional flow contributions to anti-Fourier heat transfer in rarefied cavity flow. Scientific Reports, 8(1), 13533

2018
[21]

Zhu, Roohi & Ebrahimi (2023), Computational study of rarefied gas flow and heat transfer in lid-driven cylindrical cavities. Phys. Fluids 35, 052012

2023
[22]

Zhang, L., Ma, W., Lou, Q., & Zhang, J. (2023). Simulation of rarefied gas flows using physics-informed neural network combined with discrete velocity method. Physics of Fluids , 35(7)

2023
[23]

Yao, S., Zhao, W., Wu, C., & Chen, W. (2023). Nonlinear constitutive calculation method of rarefied flow based on deep convolution neural networks. Physics of Fluids, 35(9)

2023
[24]

Huang, H., Cai, G., Wei, C., Zhang, B., Cui, X., Zhao, Y ., ... & He, B. (2025). Efficient prediction of aerodynamic forces in rarefied flow using convolutional neural network based multi- process method. Machine Learning: Science and Technology

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A., Molecular Gas Dynamics and the Direct Simulation of Gas Flows, Oxford Univ

Bird, G. A., Molecular Gas Dynamics and the Direct Simulation of Gas Flows, Oxford Univ. Press, 1994

work page 1994

[2] [2]

A., The DSMC Method, 2013

Bird, G. A., The DSMC Method, 2013

work page 2013

[3] [3]

Akhlaghi, H., Roohi, E., & Stefanov, S. (2023). A comprehensive review on micro-and nano- scale gas flow effects: Slip -jump phenomena, Knudsen paradox, thermally -driven flows, and Knudsen pumps, Physics Reports, 997, 1-60

work page 2023

[4] [4]

Variable Soft Sphere molecular model for accurate rarefied gas flow simulation,

Koura, K. & Matsumoto, H., “Variable Soft Sphere molecular model for accurate rarefied gas flow simulation,” Phys. Fluids A, 3 (10), 2459–2465 (1991)

work page 1991

[5] [5]

Collision sampling schemes for the Boltzmann equation and DSMC,

Ivanov, M. S. & Markelov, G. N., “Collision sampling schemes for the Boltzmann equation and DSMC,” USSR Comput. Math. Math. Phys., 24 (2), 84–96 (1985)

work page 1985

[6] [6]

Goshayeshi, E

B. Goshayeshi, E. Roohi, S. Stefanov, DSMC simulation of hypersonic flows using an improved SBT-TAS technique, J. Comput. Phys. 303 (2015) 28–44

work page 2015

[7] [7]

Shoja-sani, E

A. Shoja-sani, E. Roohi, S. Stefanov, Evaluation of the generalized Bernoulli trial-transient adaptive subcell ( GBT-TAS ) collision scheme in treating rarefied gas flows, Comput. Fluids. 213 (2020) 104740

work page 2020

[8] [8]

Javani, E

M. Javani, E. Roohi, S. Stefanov, Symmetrized generalized and simplified Bernoulli-trials collision schemes in DSMC, Comput. Fluids. 272 (2024) 106188

work page 2024