acceptodds
Under review as a conference paper at ICLR 2027

KF-SPINN: Separable Physics-informed Neural Networks for Kinetic-Fluid regimes

Abstract

Kinetic equations, such as the BGK model of the Boltzmann equation, describe particle dynamics across both continuum and rarefied regimes within a single partial differential equation (PDE), but the additional microscopic velocity dimension makes classical grid-based solvers computationally expensive. Physics-informed neural network (PINN)-based approaches, such as separable PINNs, offer a promising mesh-free alternative for high-dimensional PDEs. However, existing PINNs for BGK solvers struggle in the low-Knudsen-number () regime, where rapid kinetic relaxation and sharp, evolving shock layers make training challenging. In this work, we present , a framework that addresses these difficulties through complementary solution and input representations. Specifically, KF-SPINN (i) structurally encodes fast relaxation through the Chapman–Enskog-inspired solution ansatz, (ii) incorporates multiscale, time-modulated lifting features aligned with shock geometry to capture evolving shock profiles, and (iii) enhances training through conservation and entropy-transport residuals combined with a self-adaptive Charbonnier penalty. KF-SPINN improves robustness across the tested low-Kn regime and reduces mean macroscopic prediction errors by up to two orders of magnitude on 1D3V shock benchmarks relative to the strongest baseline evaluated, while keeping GPU memory requirements and training times modest.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physics-informed neural networks, kinetic equations, BGK model, Boltzmann equation, multiscale modeling, stiff PDEs, shock capturing, Chapman–Enskog expansion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EvJcOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KF-SPINN, a physics-informed neural network method for solving the BGK equation in the low-Knudsen-number regime. The method builds on SPINN-BGK and introduces three main modifications: a Chapman–Enskog-based representation with an initial-layer correction (CE-ramp), multiscale shock-aligned lifting features, and a training loss combining conservation and entropy-transport residuals with a Charbonnier penalty. The authors test the method on smooth problems and 1D3V and 2D3V Riemann problems.

Critical strengths
  1. The improvement on the 1D shock problem is substantial. The improvement becomes more pronounced as Kn decreases, which is particularly relevant since resolving sharp shocks at small Kn is the main difficulty addressed by the paper.
  2. The method makes good use of the structure of the BGK equation, and the ablations are informative. The CE-ramp representation accounts for rapid relaxation toward the hydrodynamic manifold, reducing what the remainder network needs to learn.
  3. The method achieves high accuracy without a large increase in computational cost.
Critical weaknesses
  1. The main methodological contributions could be clarified. The proposed method combines several ideas from existing work, including Chapman–Enskog expansions, shock-aligned lifting, and conservation-based training. While the combination is effective, it would be helpful if the authors could explain more clearly which components are new and how they differ from the closest related methods.
  2. The method relies on shock information whose accuracy and availability are not fully examined. In 1D, the shock speed is prescribed from the Euler Riemann solution, while in 2D, the shock surface is estimated from a preliminary PINN and then fixed for the second stage. Table 3 shows that shock alignment is essential for accuracy, but the paper does not test how sensitive the results are to errors in the prescribed shock location. The random-speed experiments in Table 15 do not address this issue because the correct shock speed is always retained. It would be useful to test the 1D problem using a predicted shock speed, as in the 2D approach, rather than the Euler Riemann speed. A small perturbation study of the shock speed and a quantitative assessment of the extracted 2D shock surface would also help clarify this dependence.
  3. The numerical results need more direct comparisons of the predicted solutions. Most results are presented as averaged relative errors, with relatively few solution profiles. I would like to see more comparisons of density, velocity, and temperature at different times and Kn values, especially near the shock and contact discontinuity. This would make it easier to judge whether the method captures the shock position, thickness, and surrounding flow correctly.
  4. The advantage over conventional BGK solvers is not clear. The paper only compares against other neural network methods. A comparison with the conservative semi-Lagrangian solver in terms of accuracy, runtime, and memory would help assess the practical value of KF-SPINN. The computational cost of obtaining the shock information should also be reported separately.
  5. The tests are limited to simple shock geometries. The 1D shock tube and radially symmetric 2D problem do not fully test the method’s ability to handle more complicated shock structures. How well would the shock extraction and lifting work for interacting shocks or irregular shock fronts?
Questions for authors

Please refer to the concerns in the weaknesses section.

6 comments
AuthorsOpenReview ↗

Response to EvJc (1/3)

Dear Reviewer EvJc,

We thank you for the positive and constructive feedback, and are glad that the reviewer found the gains on 1D shock problems substantial, especially at small Kn, CE-ramp a good use of the BGK structure, the ablations informative, and the computational overhead modest. We address each comment below; revisions are marked in blue in the manuscript.

W1. Clarification of novelty.

We agree that KF-SPINN builds on existing ideas; our contributions are BGK-specific extensions of them:

  • Relaxation-aware representation (Sec. 4.1). Micro–macro decompositions [1] leave the entire non-equilibrium part to a generic network. CE-ramp instead builds the Chapman–Enskog correction into the solution and activates it through the initial-layer factor, encoding fast relaxation as an inductive bias. The ablation applying the same ramp to the micro–macro decomposition shows that the factor alone does not yield this gain.
  • Shock-aligned lifting (Sec. 4.2). Building on lift-and-embed learning [2], TAL-PINN [3] lifts with one r-adaptive coordinate per spatial dimension, obtained from a coarse high-viscosity solution. We instead extract the shock with a Canny-type detector [4], fit it as an implicit neural level set via IGR [5], and embed it through a bank of smoothed-Heaviside scales whose contributions are modulated in time.
  • Training objective (Sec. 4.3). The conservation residual follows APNN [6], with moment-dependent weights as in SPINN-BGK. We add an entropy-transport residual, an equality obtained by combining the conservation residuals, which targets entropy transport across contact waves. We also make the transition scale of the Charbonnier penalty [7] self-adaptive to the residual magnitude rather than fixed.

Section 5.2 and Appendix G.2 ablate the representation, lifting features, and objective separately.

Revision location: Appendix G.2 (“CE and ramp”), Table 23.

W2. Two-stage method for the 1D3V Riemann problem and shock sensitivity.

Two-stage method for 1D3V Riemann. As suggested, we replace the prescribed Euler shock speed with a space–time shock level set learned from the stage-1 prediction. We follow the 2D3V schedule (50k stage-1 iterations, 20k IGR iterations, and 100k stage-2 iterations) with 384 random space–time collocation points in stage 2, retaining the original 1D3V domains, velocity quadrature, and optimizer settings:

GeometryKn=10−4\mathrm{Kn}=10^{-4}3×10−43\times10^{-4}5×10−45\times10^{-4}7×10−47\times10^{-4}10−310^{-3}
Prescribed Euler speed (Table 1)6.35e-042.22e-041.68e-041.20e-041.03e-04
Learned two-stage geometry (Table 22)9.67e-036.46e-042.60e-041.37e-047.42e-05

The learned geometry removes the need for a prescribed speed, keeping errors below 6.5×10−46.5\times10^{-4} for Kn≥3×10−4\mathrm{Kn}\geq3\times10^{-4}. Accuracy degrades at Kn=10−4\mathrm{Kn}=10^{-4}, consistent with the sensitivity analysis below.

Revision location: Appendix G.2 (“2 stage method for 1D3V Riemann” and “Accuracy and wall-clock time”), Table 22 and Figure 13.

Shock sensitivity. In 1D, we replace the shock speed c=1.84447c=1.84447 with c∈{1.5,1.7,1.8,2.0,2.2}c\in\{1.5,1.7,1.8,2.0,2.2\}; unlike Table 15, the correct speed is not retained. In 2D, we perturb the learned shock geometry. At Kn=10−4\mathrm{Kn}=10^{-4}:

ProblemUnperturbed errorError range under perturbation
1D3V (Table 18)5.77e-045.81e-03–1.62e-02
2D3V (Table 19)9.16e-037.65e-03–9.98e-03

The 1D results reveal sensitivity in the thinnest-shock case, which diminishes as Kn increases (Table 18). In 2D, the error remains stable: auxiliary networks recovered the unperturbed shock surface and reduced the error for seeds 1–4, whereas recovery failed for seed 0, whose error increased for A=−0.02A=-0.02 (Table 20). Thank you for suggesting this test; it revealed a useful compensating role of the auxiliary features.

To assess the extracted 2D shock surface, we note that the BGK shock is a layer of finite thickness, so its location is not uniquely defined. We therefore overlay the extracted surface on the reference fields (Figures 15 and 16).

Revision location: Appendix G.2 (“Shock speed sensitivity”), Tables 18–20; Appendix G.3 (“The accuracy of IGR”), Figures 15 and 16.

AuthorsOpenReview ↗

Response to EvJc (2/3)

W3. More prediction profiles

We added profile plots for the 1D3V Riemann problem: ρ\rho, uxu_x, and TT at five times for Kn=10−4\mathrm{Kn}=10^{-4} (Figure 17), and close-ups of the contact and the shock at t=0.1t=0.1 for Kn=10−4\mathrm{Kn}=10^{-4} and 10−310^{-3} in the format of Figure 1 (Figure 18). They show the three aspects you asked about:

  • Shock position. KF-SPINN places the shock at the reference position at both Kn. NSR matches the reference at Kn=10−3\mathrm{Kn}=10^{-3} but shifts the shock by about 0.006 at Kn=10−4\mathrm{Kn}=10^{-4}.
  • Shock thickness. KF-SPINN reproduces the reference thickness, whereas SPINN-BGK, Ko et al., and NSR at Kn=10−4\mathrm{Kn}=10^{-4} smear the shock.
  • Surrounding flow. KF-SPINN matches the contact and the post-shock state; SPINN-BGK and Ko et al. overshoot the temperature behind the shock.

Revision location: Appendix G.3 (“Prediction profiles”), Figures 17 and 18.

W4. Comparison with the conservative semi-Lagrangian solver and cost of shock extraction.

Comparison with the conservative semi-Lagrangian solver. We compared KF-SPINN with the conservative semi-Lagrangian (CSL) solver [8] on the 1D3V and 2D3V Riemann problems at Kn=10−4\mathrm{Kn}=10^{-4}. Errors are with respect to the Nx=40960N_x=40960 reference in 1D3V and to a new axisymmetric reference in 2D3V (see our response to Reviewers ZTjJ and SEHq). KF-SPINN runtime and memory are from Table 1 (one A100; the 2D3V time includes both stages and level-set fitting).

ProblemMethodGrid (Nx,Nv)(N_x, N_v)ErrorTime (s)Peak memory (GiB)GPU
1D3V RiemannCSL (DIRK2)51205120, 97×25297\times25^21.63e-493011.67RTX PRO 6000 ×1
CSL (DIRK2)1024010240, 97×25297\times25^22.68e-5373132.74RTX PRO 6000 ×1
CSL (DIRK2)2048020480, 97×25297\times25^23.30e-6676964.80 (total)RTX PRO 6000 ×2
KF-SPINN–6.35e-415500.94A100 ×1
2D3V RiemannCSL (BDF3)3302330^2, 21321^32.19e-271153.99 (total)RTX PRO 6000 ×2
CSL (BDF3)6602660^2, 21321^31.45e-2461258.14 (total)RTX PRO 6000 ×3
KF-SPINN–1.81e-2863 (total)1.75A100 ×1

In 2D3V, the error of KF-SPINN lies between those of the two CSL grids. The solver is faster (71 s and 461 s versus 863 s, on different GPUs) but needs 88 and 148 times the memory of KF-SPINN and two to three powerful GPUs; memory limits the solver to 6602660^2 spatial cells. The memory of the grid-based solver grows with the product of the spatial and velocity grid sizes, which is where the separable representation pays off. In 1D3V, the solver is more accurate even on the coarsest grid at lower cost, and we do not claim an advantage there. KF-SPINN also gives a continuous solution that can be evaluated at any (t,x,v)(t, x, v).

Revision location: Appendix H.3, Table 27.

Shock extraction cost. We report the runtime of each stage of the 2D3V pipeline separately (one seed):

ComponentBudgetTime
Preliminary prediction50k training iterations413.85 s
Shock detection and level-set fittingDetection + 20k IGR iterations20.47 s
Stage-2 training100k training iterations452.99 s

Shock detection and level-set fitting account for about 2% of the total runtime.

Revision location: Appendix G.2 (“Training time and sampling strategies” and “Accuracy and wall-clock time”), Table 21 and Figure 13.

W5. More complex shock structures.

1D3V shock–density-wave interaction problem. We add a 1D3V problem in which the shock interacts with a density wave, as in the Shu–Osher problem [9] but at the length scale of our 1D problems. The right state of the Sod problem is replaced by a stationary entropy wave:

ρ0=(1−s)⋅1+s⋅0.125 w(x),T0=(1−s)⋅1+s⋅0.8w(x),u0=0,\rho_0=(1-s)\cdot1+s\cdot0.125\,w(x),\quad T_0=(1-s)\cdot1+s\cdot\frac{0.8}{w(x)},\quad u_0=0,

s(x)=12(1+tanh⁡100x),w(x)=1+0.4cos⁡(2π(x−0.5)/0.1).s(x)=\tfrac12(1+\tanh100x),\quad w(x)=1+0.4\cos\big(2\pi(x-0.5)/0.1\big).

Domain, final time, boundary conditions, and training budget follow the 1D3V Sod problem, and the reference is a Chu-reduced solution at Nx=40960N_x=40960 (refinement error <10−5<10^{-5}). The shock speed varies over time as the shock passes through the density wave, so the shock geometry is not known in advance. The two-stage variant is therefore useful even in 1D3V: it extracts this time-varying shock from its stage-1 prediction and lifts along it. Mean macroscopic relative L2L^2 errors at t=0.1t=0.1 are (Table 31):

Kn\mathrm{Kn}SPINN-BGKKo et al.NSRKF-SPINN (constant speed)KF-SPINN (two-stage)
10−410^{-4}1.38e-011.06e-014.01e-021.22e-021.07e-02

Both the two-stage variant and constant Sod speed stay well below the baselines. In Figure 20, KF-SPINN follows the reference through the shock and the density wave, whereas SPINN-BGK and Ko et al. smear them.

Revision location: Appendix H.4 (“1D3V shock–density-wave interaction problem”), Eq. (55), Table 31, and Figure 20.

AuthorsOpenReview ↗

Response to EvJc (3/3)

2D3V elliptic problem. We add a non-axisymmetric 2D3V problem in which the circular interface is replaced by a 45∘45^\circ-rotated ellipse with axis ratio 2:12{:}1 and the same area. As in the circular case, the two-stage pipeline learns the evolving shock geometry from its stage-1 prediction. Mean macroscopic relative L2L^2 errors at t=0.07t=0.07 are:

Kn\mathrm{Kn}SPINN-BGKKo et al.NSRKF-SPINN
10−410^{-4}5.91e-028.97e-022.83e-021.80e-02
10−310^{-3}5.02e-027.83e-024.48e-034.80e-03

KF-SPINN attains the lowest error at Kn=10−4\mathrm{Kn}=10^{-4} and the second lowest at 10−310^{-3}, extending the evidence to asymmetric shock propagation.

Shock intersection, reflection, and splitting remain future work. We will release the code upon acceptance to support further development of neural gas-kinetic solvers.

Revision location: Appendix H.4 (“2D3V elliptic problem”), Eq. (54) and Table 30.

We hope these responses address your concerns. Thank you again for your constructive feedback, which has substantially improved the clarity and completeness of our work.

Best regards,

Submission #23105 Author Team


References:

[1] Qin Lou, Xuhui Meng, and George Em Karniadakis. Physics-informed neural networks for solving forward and inverse flow problems via the Boltzmann-BGK formulation. Journal of Computational Physics, 447:110676, 2021.

[2] Qi Sun, Zhenjiang Liu, Lili Ju, and Xuejun Xu. Lift-and-embed learning method for solving scalar hyperbolic equations with discontinuous solutions. Journal of Scientific Computing, 107(3):87, 2026.

[3] Yameng Zhu, Weibing Deng, and Ran Bi. A two-stage adaptive lifting PINN framework for solving viscous approximations to hyperbolic conservation laws. arXiv preprint arXiv:2511.04490, 2025.

[4] Takeshi R. Fujimoto, Taro Kawasaki, and Keiichi Kitamura. Canny-Edge-Detection/Rankine-Hugoniot-conditions unified shock sensor for inviscid and viscous flows. Journal of Computational Physics, 396:264–279, 2019.

[5] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pp. 3789–3799. PMLR, 2020.

[6] Shi Jin, Zheng Ma, and Keke Wu. Asymptotic-preserving neural networks for multiscale kinetic equations. Communications in Computational Physics, 35(3):693–723, 2024.

[7] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of 1st International Conference on Image Processing, volume 2, pp. 168–172. IEEE, 1994.

[8] Seung Yeon Cho, Sebastiano Boscarino, Giovanni Russo, and Seok-Bae Yun. Conservative semi-Lagrangian schemes for kinetic equations Part II: Applications. Journal of Computational Physics, 436:110281, 2021.

[9] Chi-Wang Shu and Stanley Osher. Efficient implementation of essentially non-oscillatory shock-capturing schemes, II. Journal of Computational Physics, 83(1):32–78, 1989.


Reviewer_SEHqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KF-SPINN for solving the BGK kinetic equation at low Knudsen numbers, particularly in the presence of shocks. The method builds on the separable representation of SPINN-BGK and consists of three main designs

A Chapman–Enskog-inspired CE-ramp solution representation that explicitly encodes the decay of the initial non-equilibrium distribution and the build-up of first-order non-equilibrium corrections; Multi-scale, time-modulated, shock-aligned lifting features that help the network represent moving sharp structures; A combination of macroscopic conservation residuals, an entropy transport residual, and an adaptive Charbonnier loss to mitigate residual imbalance during training. The method is evaluated on 1D shock tubes, a 2D radial Riemann problem, and several problems with smooth and non-equilibrium initial conditions. On some 1D low-Kn tests, the macroscopic error is reduced by roughly two orders of magnitude relative to the strongest existing baseline; improvements are also reported in 2D, but with smaller gains.

Critical strengths

Fast collisional relaxation and moving shocks are learning difficulties of different natures. Addressing the former through the solution representation and the latter through the input representation is more targeted than simply increasing network capacity or tuning loss weights. The CE-ramp has a sound physical motivation and good interpretability. The exponential turn-off and turn-on terms directly correspond to the decay of the initial non-equilibrium layer and the build-up of the non-equilibrium correction. The authors study the solution representation, shock alignment, multi-scale features, time modulation, and training objectives. The 2D method extracts geometry from the first-stage prediction rather than directly using the shock location from the reference solution, and the reported runtime of the full pipeline includes both training stages and the level-set construction.

Critical weaknesses

The appendix shows that KF-SPINN simultaneously adjusts time normalization, the learning-rate schedule, warmup, hard initial-condition constraints, loss weights, and the collocation scheme. In particular, the second stage in 2D replaces the original tensor-product configuration with joint spatiotemporal collocation, along with modified learning rates and Kn-dependent conservation loss weights. These changes may be reasonable engineering choices, but they make it difficult to fully disentangle the sources of the accuracy and efficiency gains. Equal iteration counts do not imply equal computational cost: in 1D, KF-SPINN is slower than SPINN-BGK, while in 2D it is faster — but the 2D speedup is accompanied by changes to the sampling and training pipeline. I suggest adding strong baselines under a unified training configuration, at least aligning the learning-rate schedule, hard initial-condition constraints, and collocation strategy, and reporting error versus actual wall-clock time. Only then can one determine how much benefit the CE-ramp, the lifting features, and the training objectives each contribute.

Questions for authors

Could the authors further verify the convergence of the 2D reference solution and report the sensitivity of the model rankings to the reference resolution? The 1D experiments use the shock speed from the corresponding Euler Riemann solution. This is a reasonable physical prior, but it is not automatically available in general kinetic problems, and it may not precisely match the full shock evolution at finite Knudsen number. Moreover, although the 2D method constructs the level set from its own prediction, it has only been validated on radially symmetric problems, and does not yet cover shock intersection, reflection, splitting, or clearly asymmetric propagation.

4 comments
AuthorsOpenReview ↗

Response to SEHq (1/3)

Dear Reviewer SEHq,

We thank the reviewer for the thoughtful and constructive feedback. We are also gratiful that he/she found our paper have a solid theoretical foundation, with compelling qualitative result. We address each comment below; revisions are marked in blue in the manuscript.

W. Unified training configuration. We agree that equal iteration counts do not isolate the effect of the training configuration. Given the computational cost, we focused the additional baseline study on SPINN-BGK, which shares our backbone and also underlies Ko et al.; NSR uses an entirely different representation. On 2D3V, we tested all 16 combinations of the learning-rate schedule, hard IC enforcement, random collocation, and time normalization (200k iterations, five seeds; see appendix below). The best error improved only modestly, from 0.05476 to 0.05117 at Kn=10−3\mathrm{Kn}=10^{-3}, and remained 0.06607 at 10−410^{-4}; several configurations diverged (NaN). Since these settings do not consistently improve SPINN-BGK, we instead ablate them within KF-SPINN, changing one setting at a time while retaining the rest:

ConfigurationKn=10−4\mathrm{Kn}=10^{-4}Kn=10−3\mathrm{Kn}=10^{-3}
Full KF-SPINN9.16e-033.22e-03
Raw time input8.65e-033.99e-03
No learning-rate decay1.18e-028.33e-03
No warmup3.53e-026.53e-03
Soft IC penalty3.29e-026.17e-03
Fixed conservation weight2.66e-023.22e-03

For the last row, λcons\lambda_{\mathrm{cons}} is fixed at 30/Tf30/T_f, its value at Kn=10−3\mathrm{Kn}=10^{-3}, so this row coincides with the full model at that Kn. Removing the learning-rate decay, warmup, hard IC, or the Kn-adaptive conservation weight increases the error at Kn=10−4\mathrm{Kn}=10^{-4}; raw time input is slightly better at 10−410^{-4} but worse at 10−310^{-3}. Over the full Kn range, the full KF-SPINN is the best or second-best configuration (Table 24). To isolate the representation itself, we also replaced CE-ramp by a micro–macro representation with the same initial-layer factor while keeping lifting, losses, and training fixed; CE-ramp attained a lower error at every Kn on the 1D3V Riemann and smooth problems (Table 23).

Revision location: Appendix G.2 (“Additional Components”, “CE-ramp for Riemann problem”, and “CE and ramp”), Tables 23–25.

W. Error versus wall-clock time. Figure 13 now reports error versus wall-clock time at Kn=10−4\mathrm{Kn}=10^{-4} for both 1D3V and 2D3V, with KF-SPINN using the learned two-stage geometry in both cases. The elapsed time includes preliminary training, level-set construction, and stage-2 training. KF-SPINN converges faster than the baselines, most clearly in 2D3V.

We also separate the effects of sampling and derivative implementation (Table 21). Jet propagation explicitly propagates and reuses the derivatives required by CE-ramp, in contrast to nested automatic differentiation (AD). Representative 2D3V results on an NVIDIA RTX A6000 (runtimes include both stages and level-set construction) are:

Space–time samplingPointsError at Kn=10−4\mathrm{Kn}=10^{-4}Nested AD (min)Jet (min)
Tensor (12,12,12)(12,12,12)17282.37e-0246.924.2
Random5129.08e-0330.013.3

Bold marks the setting used in Table 1. Random sampling attains a lower error than the tensor grid with fewer points, and jet propagation roughly halves the training time.

Revision location: Appendix G.2 (“Training time and sampling strategies” and “Accuracy and wall-clock time”), Table 21 and Figure 13.

AuthorsOpenReview ↗

Response to SEHq (2/3)

Q. 2D3V reference accuracy. We re-solved the 2D3V Riemann problem with an axisymmetric solver [1] combined with the Chu reduction [2], which applies because the problem is rotationally symmetric. Refining the radial grid up to Nr=19200N_r=19200 at fixed CFL, which also refines the time step, gives an estimated error of 55–6×10−46\times10^{-4} for the finest solution, an order of magnitude below the model errors. Mean macroscopic relative L2L^2 errors of the saved Table 1 checkpoints against each reference are (Kn=10−4\mathrm{Kn}=10^{-4}, five seeds):

MethodOriginal reference (6602660^2)Axisymmetric reference
SPINN-BGK5.30e-25.71e-2
Ko et al.5.07e-25.61e-2
NSR2.21e-23.08e-2
KF-SPINN9.16e-31.81e-2

The ranking is the same for both references. All errors increase because the original reference differs from the finest solution by about 1.2×10−21.2\times10^{-2}, and KF-SPINN remains the most accurate.

Note on reported values. Some entries differ slightly between this response and the manuscript (e.g., the original-reference column above and Table 1: 5.30×10−25.30\times10^{-2} vs. 5.05×10−25.05\times10^{-2} for SPINN-BGK). These small differences arise from running experiments on different GPU hardware. In the camera-ready version, we will rerun all experiments on a single hardware configuration so that all reported values are consistent.

Revision location: Appendix H.4 (“Convergence of the axisymmetric reference” and “2D3V Riemann problem”), Tables 28 and 29.

Q. Two-stage method for 1D3V Riemann. We now replace the prescribed Euler shock speed with a space–time shock level set learned from the stage-1 prediction. We follow the 2D3V schedule (50k stage-1 iterations, 20k IGR iterations, and 100k stage-2 iterations) with 384 random space–time collocation points in stage 2, retaining the original 1D3V domains, velocity quadrature, and optimizer settings:

GeometryKn=10−4\mathrm{Kn}=10^{-4}3×10−43\times10^{-4}5×10−45\times10^{-4}7×10−47\times10^{-4}10−310^{-3}
Prescribed Euler speed (Table 1)6.35e-042.22e-041.68e-041.20e-041.03e-04
Learned two-stage geometry (Table 22)9.67e-036.46e-042.60e-041.37e-047.42e-05

The learned geometry removes the need for a prescribed speed, keeping errors below 6.5×10−46.5\times10^{-4} for Kn≥3×10−4\mathrm{Kn}\geq3\times10^{-4}. Accuracy degrades at Kn=10−4\mathrm{Kn}=10^{-4}, consistent with the shock-speed sensitivity observed at this Kn (Table 18).

Revision location: Appendix G.2 (“Shock speed sensitivity”, “2 stage method for 1D3V Riemann”, and “Accuracy and wall-clock time”), Tables 18 and 22, and Figure 13.

Q. Geometries beyond radial symmetry. 1D3V shock–density-wave interaction problem. We add a 1D3V problem in which the Euler shock speed no longer describes the shock evolution: the Sod shock runs into a stationary entropy wave (density amplitude 40%, wavelength 0.1, uniform pressure) placed ahead of it, as in the Shu–Osher problem but at the length scale of our 1D problems. Domain, final time, boundary conditions, and the training budget of every method follow the 1D3V Sod problem; the reference is a Chu-reduced solution at Nx=40960N_x=40960 (refinement error <10−5<10^{-5}). At Kn=10−4\mathrm{Kn}=10^{-4}, the shock speed of the BGK reference varies over time as the shock passes through the density wave, so the shock trajectory is no longer straight. We keep the constant Sod shock speed in KF-SPINN, which is now incorrect, and also run the two-stage method, which needs no prescribed speed. Mean macroscopic relative L2L^2 errors at t=0.1t=0.1 are (Table 31; mean over seeds 0–4):

Kn\mathrm{Kn}SPINN-BGKKo et al.NSRKF-SPINN (constant speed)KF-SPINN (two-stage)
10−410^{-4}1.38e-011.06e-014.01e-021.22e-021.07e-02

With an incorrect prescribed speed, KF-SPINN still improves on SPINN-BGK and Ko et al. by a factor of 9–11 and on NSR by about 2.5. The learned geometry gives a comparable error, so the gains here do not rely on the Euler speed being exact. Profiles are shown in Figure 20.

Revision location: Appendix H.4 (“1D3V shock–density-wave interaction problem”), Eq. (55), Table 31, and Figure 20.

AuthorsOpenReview ↗

Response to SEHq (3/3)

2D3V elliptic problem. We add a non-axisymmetric 2D3V problem in which the circular interface is replaced by a 45∘45^\circ-rotated ellipse with axis ratio 2:12{:}1 and the same area. As in the circular case, the two-stage pipeline learns the evolving shock geometry from its stage-1 prediction. Mean macroscopic relative L2L^2 errors at t=0.07t=0.07 are:

Kn\mathrm{Kn}SPINN-BGKKo et al.NSRKF-SPINN
10−410^{-4}5.91e-028.97e-022.83e-021.80e-02
10−310^{-3}5.02e-027.83e-024.48e-034.80e-03

KF-SPINN attains the lowest error at Kn=10−4\mathrm{Kn}=10^{-4} and the second lowest at 10−310^{-3}, extending the evidence to asymmetric shock propagation.

Shock intersection, reflection, and splitting remain future work. We will release the code upon acceptance to support further development of neural gas-kinetic solvers.

Revision location: Appendix H.4 (“2D3V elliptic problem”), Eq. (54) and Table 30.

We hope these responses address your concerns. Thank you again for your constructive feedback, which has helped us make the comparisons more controlled and transparent.

Best regards,

Submission #23105 Author Team


References:

[1] Luc Mieussens. Discrete-velocity models and numerical schemes for the Boltzmann-BGK equation in plane and axisymmetric geometries. Journal of Computational Physics, 162(2):429–466, 2000.

[2] C. K. Chu. Kinetic-theoretic description of the formation of a shock wave. The Physics of Fluids, 8(1):12–22, 1965.


Appendix : 2D3V Riemann, SPINN-BGK, 200k iteration

LR changeHard ICRandomTime normKn=1e-3Kn=1e-4
OFFOFFOFFOFF0.054760.06607
OFFOFFOFFON0.051170.06925
OFFOFFONOFF0.062390.07095
OFFOFFONON0.058900.06968
OFFONOFFOFF0.153210.11915
OFFONOFFON0.168960.14004
OFFONONOFF0.108040.12212
OFFONONON0.121070.13049
ONOFFOFFOFFNaN (0/5)NaN (0/5)
ONOFFOFFONNaN (0/5)NaN (0/5)
ONOFFONOFF0.155500.18933
ONOFFONON0.160250.18632
ONONOFFOFFNaN (0/5)NaN (0/5)
ONONOFFONNaN (0/5)NaN (0/5)
ONONONOFF0.179090.20215
ONONONON0.184680.18741
AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer SEHq,

Thank you again for your careful review and helpful comments. As the discussion period ends in about two days, we would appreciate your feedback on whether our responses and the added experiments in Appendices G.2 and H.4 have addressed your concerns regarding comparison fairness, reference accuracy, and broader shock dynamics. We welcome any further questions and would be glad to clarify them before the discussion closes.

Best regards,

Submission #23105 Author Team

Reviewer_ZTjJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes KF-SPINN, a separable physics-informed neural network for solving the BGK kinetic equation, focusing on the temporal stiffness and sharp shock layers encountered at small Knudsen numbers. Building on SPINN-BGK, it combines three components: (i) a Chapman–Enskog-inspired “CE-ramp” solution representation that explicitly incorporates fast initial relaxation and gradually activates a first-order nonequilibrium correction; (ii) multiscale, time-modulated lifting features aligned with shock geometry; and (iii) a training objective combining BGK, macroscopic conservation, and entropy-transport residuals with a self-adaptive Charbonnier penalty. For two-dimensional problems, shock geometry is estimated through a preliminary neural solution and a learned level-set representation.

Validation compares KF-SPINN with SPINN-BGK, a velocity-weighted variant, and NSR against conservative semi-Lagrangian reference solutions. Experiments cover 1D3V Sod shock tubes and a radially symmetric 2D3V Riemann problem at Knudsen numbers from 10−410^{-4} to 10−310^{-3}, with additional smooth, nonequilibrium-initialization, and higher-Knudsen-number tests. Across five runs, the reported mean macroscopic errors are lowest among the evaluated methods on all main benchmarks, with improvements reaching approximately two orders of magnitude in 1D3V. Ablations examine the individual representations, lifting features, and loss components, and the paper reports training time and memory usage. The evidence is strongest for the tested benchmark geometries; the authors acknowledge that the 2D3V reference discretization error has not been established to be substantially below the reported model errors.

Critical strengths
  1. Strong empirical performance in the tested low-Knudsen-number regime. KF-SPINN consistently improves upon the evaluated neural baselines on the 1D3V shock benchmarks, with mean macroscopic errors below 10−310^{-3} and gains reaching approximately two orders of magnitude. Results averaged over five runs and supported by reference-grid refinement checks provide convincing evidence of practical value on these benchmarks.

  2. Physically motivated design with informative ablations. The CE-ramp representation and shock-aligned lifting address distinct difficulties: fast kinetic relaxation and sharp, evolving shock layers. Targeted ablations substantiate the benefits of these components and the conservation-aware training objective, providing useful guidance for designing neural kinetic solvers.

Critical weaknesses
  1. The methodological novelty is incremental, and its broader significance is not sufficiently demonstrated. CE-ramp is a physically motivated extension of existing micro–macro and Chapman–Enskog representations, while the lifting features and training objective build on established techniques. The ablations support the effectiveness of the combination, but validation is restricted to a one-dimensional shock tube and a radially symmetric two-dimensional Riemann problem. Because shock alignment is central to the gains, these experiments leave substantial uncertainty about whether the method offers a broadly useful advance beyond the selected geometries.

  2. The reference accuracy limits the conclusions from the higher-dimensional benchmark. At Kn = 0.0001, the reported 2D3V reference-grid refinement difference is 0.00919, comparable to KF-SPINN’s reported error of 0.00916. This difference is not itself an estimate of the remaining reference error, but it shows that reference convergence has not been established at the accuracy needed to substantiate the reported model errors. Consequently, the accuracy ranking against the underlying PDE solution remains uncertain, weakening the principal evidence for the method’s higher-dimensional effectiveness.

Questions for authors
  1. Clarification on reference accuracy. Can you provide further convergence evidence for the 2D3V reference solutions, including spatial, temporal, and velocity discretization? In particular, does the ranking of the methods remain consistent across the available reference resolutions, given that the reported spatial refinement differences are comparable to KF-SPINN’s errors?

  2. Clarification on sensitivity to shock geometry. How much does accuracy deteriorate when the supplied 1D shock speed is perturbed or the preliminary 2D solution yields a less accurate shock level set? A quantitative sensitivity analysis on the existing benchmarks would materially clarify whether the gains require precise geometric information.

  3. Clarification on the contribution beyond existing representations. What advantage does CE-ramp provide over a micro–macro representation equipped with the same explicit initial-layer decay, while keeping lifting, losses, and training settings fixed? This comparison would help isolate the value of gradually activating the CE correction from that of directly encoding fast relaxation.

3 comments
AuthorsOpenReview ↗

Response to ZTjJ (1/2)

Dear Reviewer ZTjJ,

We thank the reviewer for the thoughtful and constructive feedback. We are also gratiful that he/she found our paper have a solid theoretical foundation, with compelling qualitative result. We address each question below; revisions are marked in blue in the manuscript.

Q1. Clarification on reference accuracy.

We agree that the refinement difference of the original 2D3V reference (9.19×10−39.19\times10^{-3}, 3302→6602330^2\to660^2) is not an estimate of its error, and that it is as large as the KF-SPINN error at Kn=10−4\mathrm{Kn}=10^{-4}. Since the problem is rotationally symmetric, we re-solved it with an axisymmetric discrete-velocity solver [1] combined with the Chu reduction [2, 3], which reduces space to the radius and removes vzv_z, so that the grids can be refined far beyond what the Cartesian solver allows. At Kn=10−4\mathrm{Kn}=10^{-4}:

  • Space and time. Successive differences decrease from 4.1×10−34.1\times10^{-3} to 3.3×10−43.3\times10^{-4} as NrN_r is refined from 1200 to 19200. The CFL number is fixed, so the time step is refined together with the radial grid.
  • Velocity. Refining each of the two velocity grids changes the solution by 22–3×10−43\times10^{-4}.

The finest solution thus has an estimated error of 55–6×10−46\times10^{-4}, an order of magnitude below the model errors. Re-evaluating the saved Table 1 checkpoints against it:

MethodOriginal reference (6602660^2)Axisymmetric reference
SPINN-BGK5.30e-25.71e-2
Ko et al.5.07e-25.61e-2
NSR2.21e-23.08e-2
KF-SPINN9.16e-31.81e-2

The ranking is unchanged. Against an intermediate reference, a finer Cartesian solution with the Chu reduction (160021600^2), KF-SPINN and NSR give 1.37×10−21.37\times10^{-2} and 2.74×10−22.74\times10^{-2}, again in the same order. All errors increase because the original reference differs from the finest solution by about 1.2×10−21.2\times10^{-2}; KF-SPINN remains the most accurate, 1.7 times below NSR. The axisymmetric solver applies only to rotationally symmetric problems.

Note on reported values. Some entries differ slightly between this response and the manuscript (e.g., the original-reference column above and Table 1: 5.30×10−25.30\times10^{-2} vs. 5.05×10−25.05\times10^{-2} for SPINN-BGK). These small differences arise from running experiments on different GPU hardware. In the camera-ready version, we will rerun all experiments on a single hardware configuration so that all reported values are consistent.

Revision location: Appendix H.4 (“Convergence of the axisymmetric reference” and “2D3V Riemann problem”), Tables 28 and 29.

Q2. Clarification on sensitivity to shock geometry.

1D3V. We perturb the shock speed cc in ψshock(t,x)=x−(xd+ct)\psi_{\mathrm{shock}}(t,x)=x-(x_d+ct); the correct speed is not retained. Mean macroscopic relative L2L^2 errors are (Table 18):

Kn\mathrm{Kn}c=1.84447c=1.84447 (baseline)c=1.5c=1.5c=1.7c=1.7c=1.8c=1.8c=2.0c=2.0c=2.2c=2.2
10−410^{-4}5.77e-047.88e-031.62e-025.81e-036.02e-037.55e-03
10−310^{-3}8.52e-059.80e-051.08e-049.32e-059.87e-051.12e-04

Sensitivity is pronounced at Kn=10−4\mathrm{Kn}=10^{-4}, where the shock layer is thinnest, and minor at 10−310^{-3}. When the shock speed is not supplied but learned with the two-stage method, errors remain below 6.5×10−46.5\times10^{-4} for Kn≥3×10−4\mathrm{Kn}\geq3\times10^{-4} but reach 9.67×10−39.67\times10^{-3} at 10−410^{-4} (Table 22).

2D3V. We perturb the learned shock geometry with perturbation parameter AA (A=0A=0 denotes the baseline). Mean macroscopic relative L2L^2 errors are (Table 19):

Kn\mathrm{Kn}A=0A=0A=−0.02A=-0.02A=−0.01A=-0.01A=−0.005A=-0.005A=+0.005A=+0.005A=+0.01A=+0.01
10−410^{-4}9.16e-038.99e-039.98e-039.96e-038.73e-037.65e-03
10−310^{-3}3.22e-033.71e-033.66e-034.26e-034.31e-034.21e-03

The errors remain relatively stable. For A=−0.02A=-0.02, the auxiliary networks recovered the unperturbed IGR surface and reduced the error for seeds 1–4, whereas recovery failed for seed 0, whose error increased (Table 20). Thus, accurate shock geometry matters most in the thinnest-shock 1D case, while in 2D the auxiliary features can compensate for geometric perturbations.

Revision location: Appendix G.2 (“Shock speed sensitivity” and “2 stage method for 1D3V Riemann”), Tables 18–20 and 22.

Beyond radial symmetry. We add a non-axisymmetric 2D3V problem with a 45∘45^\circ-rotated elliptical interface (axis ratio 2:12{:}1). Mean macroscopic relative L2L^2 errors at t=0.07t=0.07 are (Table 30):

Kn\mathrm{Kn}SPINN-BGKKo et al.NSRKF-SPINN
10−410^{-4}5.91e-028.97e-022.83e-021.80e-02
10−310^{-3}5.02e-027.83e-024.48e-034.80e-03

KF-SPINN attains the lowest error at Kn=10−4\mathrm{Kn}=10^{-4} and the second lowest at 10−310^{-3}.

AuthorsOpenReview ↗

Response to ZTjJ (2/2)

1D3V shock–density-wave interaction problem. As in the Shu–Osher problem, the shock of the 1D3V Riemann problem runs into a stationary entropy wave (density amplitude 40%, wavelength 0.1) placed ahead of it, with the domain, final time, boundary conditions, and training budget of the original problem. At Kn=10−4\mathrm{Kn}=10^{-4}, the shock speed of the reference varies over time as the shock passes through the density wave, so the shock trajectory is no longer straight. The constant speed supplied to KF-SPINN is therefore incorrect throughout, which directly tests whether the gains require precise geometric information. Mean macroscopic relative L2L^2 errors at t=0.1t=0.1 are (Table 31; mean over seeds 0–4):

Kn\mathrm{Kn}SPINN-BGKKo et al.NSRKF-SPINN (constant speed)KF-SPINN (two-stage)
10−410^{-4}1.38e-011.06e-014.01e-021.22e-021.07e-02

Even with the incorrect speed, KF-SPINN remains about 2.5 times as accurate as NSR and an order of magnitude more accurate than SPINN-BGK and Ko et al. The two-stage variant, which extracts the time-varying shock from its own stage-1 prediction, is comparable. Together with the ellipse, this extends the evaluation to asymmetric and time-varying shock propagation; shock intersection, reflection, and splitting remain future work.

Revision location: Appendix H.4 (“2D3V elliptic problem” and “1D3V shock–density-wave interaction problem”), Eqs. (54) and (55), Tables 30 and 31, and Figure 20.

Q3. Clarification on the contribution beyond existing representations.

We ran exactly this comparison, adding the factor 1−e−t/Kn1-e^{-t/\mathrm{Kn}} to the remainder part of the micro–macro representation while keeping lifting, losses, and training (including the hard initial condition) fixed. On the 1D3V Riemann and smooth problems (five Kn), the factor alone did not consistently reduce the error, whereas CE-ramp was lower, so the gain comes from gradually activating the CE correction.

Revision location: Appendix G.2 (“CE and ramp”), Table 23.

We hope these responses address your concerns. Thank you again for your constructive feedback, which has helped us strengthen the evidence for our method.

Best regards,

Submission #23105 Author Team


References:

[1] Luc Mieussens. Discrete-velocity models and numerical schemes for the Boltzmann-BGK equation in plane and axisymmetric geometries. Journal of Computational Physics, 162(2):429–466, 2000.

[2] C. K. Chu. Kinetic-theoretic description of the formation of a shock wave. The Physics of Fluids, 8(1):12–22, 1965.

[3] Sudarshan Tiwari, Axel Klar, and Giovanni Russo. A meshfree arbitrary Lagrangian–Eulerian method for the BGK model of the Boltzmann equation with moving boundaries. Journal of Computational Physics, 458:111088, 2022.


AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer ZTjJ,

Thank you again for your valuable time and constructive feedback. With only about two days remaining in the discussion period, we wanted to gently follow up on whether our responses and the additional results in Appendices G.2 and H.4 have adequately addressed your concerns about reference accuracy, shock-geometry sensitivity, and the CE-ramp contribution. Please let us know if any further clarification would be helpful; we would be happy to discuss any remaining concerns before the window closes.

Best regards,

Submission #23105 Author Team

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Responses and Revisions

Dear Area Chair,

Thank you for handling our submission. We have posted our detailed responses and revised manuscript. To assist your assessment, we summarize the main additions:

Reference accuracy. We substantially refined the 2D3V reference and added spatial, temporal, and velocity convergence studies. KF-SPINN retains the lowest error among the evaluated methods against the refined reference (Appendix H.4).

CE-ramp contribution. Under otherwise identical settings, CE-ramp outperforms the micro–macro baseline equipped with the same initial-layer factor, supporting the benefit of the CE correction (Appendix G.2).

Shock geometry and generality. We quantified sensitivity to shock geometry, evaluated learned 1D geometry without a prescribed speed, and added non-axisymmetric and shock–density-wave interaction problems (Appendices G.2 and H.4).

Comparison fairness and efficiency. We added controlled training-setting studies, error-versus-wall-clock-time curves, and accuracy, runtime, and memory comparisons with the numerical solver (Appendices G.2 and H.3).

We believe these additions substantially strengthen the empirical support for our conclusions and would particularly appreciate your assessment of how this additional evidence addresses the concerns raised in the reviews. Revisions are marked in blue, and we remain available for any further clarification.

Best regards,

Submission #23105 Author Team

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.