acceptodds
Under review as a conference paper at ICLR 2027

AnisoLift: Anisotropic Latent Representations for Coarse Particle Liquid Enhancement

Abstract

Particle-based liquid simulation is widely used in graphics and physical modeling, but high-resolution rollouts remain computationally expensive. Consequently, many methods aim to recover fine-scale dynamics and dense transport patterns from coarse particle simulations. However, these methods typically rely on additional particle generation, which still incurs considerable computational overhead and leads to poor representation. To this end, we propose AnisoLift, a structured latent closure framework that augments each coarse particle with learnable anisotropic ellipsoidal components. This allows the model to capture directional local structure from the underlying high-resolution flow without introducing extra particles. Given a coarse simulation, our model predicts residual corrections to particle states to bring the updated state closer to the aligned high-resolution teacher. Our training objective jointly supervises particle dynamics and anisotropic geometric structure, encouraging consistency with the reference particle states and local geometric structure. Experiments show that our approach enhances coarse liquid simulations through improving fidelity to fully resolved flow behavior.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for science, Physical Intelligence, Fluid dynamics, simulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fwznOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AnisoLift enhances precomputed coarse particle-liquid trajectories while keeping the particle count fixed. It predicts per-particle position and velocity corrections and attaches an anisotropic covariance to represent local structure learned from paired high-resolution simulations. The corrections are applied to each saved frame and are not fed back into the solver. Results cover SPH and MPS simulations in both 2D and 3D, with position, velocity, kinetic-energy, and limited free-surface evaluations.

Critical strengths
  1. The fixed-particle setting is useful. The method improves coarse states without adding particles or changing the solver, and the results cover several 2D and 3D SPH/MPS cases. This makes the representation idea relevant where a full high-resolution rollout is too costly.

  2. The surface diagnostics add evidence beyond particle-wise errors. On 2D dam break, the reported occupancy IoU, F1, and surface-height error improve for both SPH and MPS. The anisotropic geometry also gives a plausible way to encode local directional structure at fixed particle count.

Critical weaknesses
  1. Much of the quantitative evaluation remains close to the constructed high-resolution supervision targets. The paper does include independent-form surface diagnostics, but only for 2D dam break. It is hard to judge from the current experiments whether the corrections improve broader physical behavior, particularly temporal consistency and conservation across the sequence. This limits how far I can interpret the particle-wise gains.

  2. The method corrects each precomputed frame independently and does not feed its output back to the solver. The paper acknowledges that it cannot repair a global transport mode missing from the coarse trajectory. This is a reasonable scope choice, but it narrows the claim of liquid enhancement. The current results do not quantify how stable the corrected states and anisotropic footprints are between adjacent frames.

Questions for authors
  1. Could the authors report a sequence-level measure of temporal variation for corrected positions, velocities, and covariance orientations, alongside the coarse and high-resolution trajectories? A short visualization alone would not settle this concern.

  2. Can the surface-level evaluation be extended beyond 2D dam break, or can the authors report frame-to-frame changes in momentum and kinetic energy relative to the high-resolution trajectories?

  3. In the runtime table, does the stated inference time include feature extraction? The text says feature extraction dominates, while the table describes an MLP forward-pass timing. Please give end-to-end post-processing cost if available.

5 comments

Q1 and W2.1. Stability of corrected states and ellipsoids between adjacent frames.

We sincerely thank the reviewer for highlighting the important issue of temporal stability in AnisoLift, which prompted us to conduct additional experiments. We provided a sequence-level measure of temporal variation for the corrected positions, velocities, and covariance orientations, alongside both coarse and high-resolution reference trajectories, as shown in the Table R3-1 below. These results were obtained on SPH simulation data for 2D DAM, 2D TGV, and 3D TGV. Apart from the MSExMSE_x and MSEvMSE_v metrics already employed in our paper, four further metrics were introduced to quantify the temporal variation of position, velocity, ellipsoid rotation, and covariance orientations. These indicators are calculated as follows, Between consecutive frames, each column’s own particles give the second differences δ2x=∥xt+1−2xt+xt−1∥2\delta^2 x = \lVert x_{t+1}-2x_t+x_{t-1}\rVert_2 (the same to δ2v\delta^2 v), the log-Euclidean change dLE(C)=∥log⁡Ct+1−log⁡Ct∥Fd_{\mathrm{LE}}(C)=\lVert \log C_{t+1}-\log C_t\rVert_F, and the rotation ΔθC=arccos⁡∣ut+1⊤ut∣\Delta\theta_C=\arccos\lvert u_{t+1}^{\top}u_t\rvert, with uu the unit major eigenvector of CC. Since the "Reference" represents the ground truth, its MSExMSE_x and MSEvMSE_v values are denoted by "-". Furthermore, as neither the "Coarse" nor the "Reference" results contain ellipsoidal particles, orientation-related metrics are not applicable, these entries are marked as "n/a." These results indicate that, compared to the "Coarse", AnisoLift’s temporal variations are closer to those of the "Reference." We highlight the manuscript change in blue in the new section D of the revised appendix.

W2.2. Support to our claim of liquid enhancement.

We particularly appreciate the reviewer’s recognition that our frame-wise refinement is a reasonable scope choice. At the same time, we acknowledge the valid concern raised in W2 regarding the scope of our claims about liquid enhancement. We thus implemented a centered smoothing approach applying a 3 or 9-frame moving average to the outputs of AnisoLift to support our claim (Ours+MA3, Ours+MA9 in the Table R3-1 below). Our results indicate that this simple method brings the magnitude of position and velocity temporal variation closer to the reference. In future work, we intend to develop a more robust mechanism to ensure positive outcomes across a broader range of scenarios as mentioned in Section 5.

Table R3-1. Sequence-level measure of temporal variation for corrected states.

QuantityCoarseOurs + MA3Ours + MA9OursReference
2D DAM
MSE$_x$ ↓ (10−310^{-3})2.5601.0751.0621.100-
MSE$_v$ ↓ (10−210^{-2})3.5922.1821.9482.234-
δ2x(≈Ref.,10−4\delta^2 x (≈ Ref.,10^{-4})4.784.864.754.874.72
δ2v(≈Ref.,10−2\delta^2 v (≈ Ref.,10^{-2})2.232.091.952.151.74
dLE(C)(≈Ref.)d_{\mathrm{LE}}(C)(≈ Ref.)n/a0.0970.0440.188n/a
ΔθC(°)(≈Ref.)\Delta\theta_C(°)(≈ Ref.)n/a1.740.851.53n/a
2D TGV
MSE$_x$ ↓ (10−510^{-5})7.7501.8871.2302.010-
MSE$_v$ ↓ (10−310^{-3})2.5652.3191.6152.216-
δ2x(≈Ref.,10−4\delta^2 x(≈ Ref.,10^{-4})7.466.785.326.725.62
δ2v(≈Ref.,10−2\delta^2 v(≈ Ref.,10^{-2})3.022.041.652.321.70
dLE(C)(≈Ref.)d_{\mathrm{LE}}(C)(≈ Ref.)n/a0.0190.0100.061n/a
ΔθC(°)(≈Ref.)\Delta\theta_C (°)(≈ Ref.)n/a1.170.921.51n/a
3D TGV
MSE$_x$ ↓ (10−510^{-5})5.192.182.145.17-
MSE$_v$ ↓ (10−310^{-3})21.267.934.4312.54-
δ2x(≈Ref.,10−2\delta^2 x(≈ Ref.,10^{-2})3.502.561.622.571.56
δ2v(≈Ref.,10−2\delta^2 v(≈ Ref.,10^{-2})14.99.378.149.087.76
dLE(C)(≈Ref.)d_{\mathrm{LE}}(C)(≈ Ref.)n/a0.0270.0180.044n/a
ΔθC(°)(≈Ref.)\Delta\theta_C(°)(≈ Ref.)n/a3.551.878.76n/a

Q2.1 Surface metrics. This requires a free surface. While in TGV and RPF the fluid fills a periodic box and in LDC it fills a closed cavity, so none of them has a free surface, the occupancy is full (IoU = 1 for every method) and the height profile is undefined. Those scenes are not proper to be evaluated at surface level. The change is in the fourth paragraph of section 4.4 in our manuscript highlighted in blue.

Q2.2. Frame-to-frame changes in momentum and kinetic energy relative to the high-resolution trajectories. We thank the reviewer for highlighting this important aspect of physical consistency. This comment led us to evaluate conservation properties according to the underlying physics of each simulation scenario. Following this suggestion, we have extended our evaluation to explicitly examine momentum and kinetic energy evolution across three scenarios, and show results in the Table R3-2 below. We highlight the corresponding manuscript change in blue in the section E of the appendix in the revision.

Specifically, kinetic energy is not expected to remain constant in any of the three scenarios, since it changes due to gravity and viscous dissipation in 2D DAM, whereas in both 2D and 3D TGV scenarios, it exhibits an overall decay due to viscous dissipation. We therefore compare its temporal evolution and frame-to-frame changes with those of the high-resolution reference. In contrast, total fluid momentum is theoretically conserved in the periodic and force-free TGV settings (up to numerical errors), and we evaluate the magnitude of its frame-to-frame change, which should vanish. In 2D DAM, where gravity and wall interactions exert external forces on the fluid, we instead compare its temporal evolution with that of the high-resolution reference.

Coarse and our corrected trajectories, we compute fluid specific kinetic energy ete_t and total momentum Pt\mathbf P_t per frame, take increments Δet=et+1−et\Delta e_t = e_{t+1}-e_t, et=12N∑i∥vi,t∥22e_t = \frac{1}{2N} \sum_i\Vert v_{i,t}\Vert_2^2; ΔPt=Pt+1−Pt\Delta \mathbf P_t = \mathbf P_{t+1}-\mathbf P_t, Pt=m∑ivi,t\mathbf P_t=m\sum_i v_{i,t} and report Mean ∣Δet∣\lvert\Delta e_t\rvert: Et∣Δet−Δetr∣E_t|\Delta e_t-\Delta e_t^{\rm r}| (likewise Mean ∥ΔPt∥\lVert\Delta \mathbf P_t\rVert), Rel. error of $ \Delta e_t = \frac{\sum_t \vert \Delta e_t - \Delta e_t^r \vert }{\sum_t \vert \Delta e_t^r \vert} $, $ \Delta \mathbf P_t = \frac{ \sum_t \Vert \Delta \mathbf P_t - \Delta \mathbf P_t^r \Vert_2 }{\sum_t \Vert \Delta \mathbf P_t^r \Vert_2} $ in the Table R3-2 below. Mean ∣Δet∣\lvert\Delta e_t\rvert and Mean ∥ΔPt∥\lVert\Delta\mathbf P_t\rVert measure how large the per-frame change. The two Rel. error columns measure how far the per-frame changes are from those of the reference, in units of the reference’s own change, lower is better. Each trajectory is computed on its own particles, values are means over frames, then mean ± std over test trajectories.

Note that “—” appears only in the Reference rows of the Rel. error columns. A relative error of the reference to itself is zero by definition and carries no information, while “n/a” marks a quantity that is undefined in that scene. In the periodic TGV box the reference changes total momentum only at round-off level, so dividing by it gives a meaningless number. We therefore leave Rel. error of ΔPt\Delta\mathbf P_t undefined there and assess conservation from Mean ∥ΔPt∥\lVert\Delta\mathbf P_t\rVert instead. As shown by the numerical results in the table below, compared to coarse particles, AnisoLift yields trajectories whose Mean ∣Δet∣(1e−3)\lvert\Delta e_t\rvert(1e-3) and ∥ΔPt∥\lVert\Delta \mathbf{P}_t\rVert are closer to those of high-resolution reference particles, meanwhile, AnisoLift's trajectories enjoy lower relative error of Δet\Delta e_t and ΔPt\Delta \mathbf{P}_t than coarse trajectories.

Table R3-2. Frame-to-frame changes in momentum and kinetic energy relative to the high-resolution trajectories.

SceneMethodMean ∣Δet∣(1e−3)\lvert\Delta e_t\rvert(1e-3)Rel. error of Δet\Delta e_t ↓Mean ∥ΔPt∥\lVert\Delta \mathbf{P}_t\rVertRel. error of ΔPt\Delta \mathbf{P}_t ↓
2D Dam breakReference1.143±0.0181.143\pm 0.018—0.0133±0.032e−20.0133\pm 0.032e-2—
2D Dam breakCoarse1.270±0.0421.270\pm0.0420.179±0.0500.179\pm 0.0500.0143±0.067e−20.0143\pm 0.067e-2$0.940\pm 0.054 $
2D Dam breakOurs1.158±0.0151.158\pm 0.015$0.136\pm 0.039 $0.0136±0.049e−20.0136\pm 0.049e-2$0.899\pm 0.043 $
2D TGVReference2.000std<10−62.000\text{std} < 10^{-6}—8.100e−11±1.4e−118.100e-11\pm 1.4e-11n/a
2D TGVCoarse1.903±0.0301.903\pm 0.0300.193±0.1090.193\pm 0.1091.701e−10±0.3e−101.701e-10\pm 0.3e-10n/a
2D TGVOurs2.006±0.0232.006\pm0.0230.172±0.1030.172\pm 0.1039.800e−11±0.5e−119.800e-11\pm 0.5e-11n/a
3D TGVReference2.081std<10−62.081\text{std} < 10^{-6}—1.10e−8±0.2e−81.10e-8\pm 0.2e-8n/a
3D TGVCoarse1.904±0.0221.904\pm 0.0220.181±0.0440.181\pm 0.0441.72e−8±0.18e−81.72e-8\pm 0.18e-8n/a
3D TGVOurs2.080±0.0492.080\pm 0.049$0.170\pm 0.099 $1.41e−8±0.7e−81.41e-8\pm 0.7e-8n/a

W1. Quantitative evaluation remains close to the constructed high-resolution supervision targets. Like learned Lagrangian simulators such as GNS or Neural SPH, we supervise with high-resolution simulation data. Unlike them, coarse and high-resolution particles do not correspond one-to-one, so each coarse particle’s target is a projection of the reference onto it, i.e., a kernel average of its high-resolution neighbors, in the spirit of Yu & Turk (2013). Table 1 measures how well this projection is recovered.

Q3. Does the inference time include feature extraction? No, Table 8 timed only the MLP forward pass. We now time the full processing of one frame, from the coarse states to corrected positions, velocities and covariances in the Table R3-3 below. Post-processing time spent is less than one single solver step in that frame on both scenes. We highlight this manuscript change in blue in the section B.2 of the revision.

Table R3-3. Full processing time of one frame.

2D DAM3D TGV
Number of Coarse/High-resolution particles1,622/5,7401,000/8,000
Complete feature extraction420 μ\mus400 μ\mus
MLP forward pass incl. assembly of C=LL⊤C = LL^\top5.8 μ\mus6.3 μ\mus
Apply corrections, copy results back160$\mu$s160$\mu$s
AnisoLift total585.8 μ\mus566.3 μ\mus
Coarse solver1,680 ms593 ms
High-resolution solver2,023 ms2,482 ms
Post-processing over the coarse solver0.035%0.095%
Post-processing over the high-resolution solver0.029%0.023%
Reviewer_Gy7zOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presented AnisoLift, which is a post-simulation augment framework that recovers fine-scale particle states from coarse particle liquid simulations. AnisoLift focuses on fixed particle-count setting, leveraging local reference neighborhood weighted sum and covariance ellipsoid to align coarse simulation to a simultaneous high-resolution state. The paper reports experimental results on various 2D/3D datasets, including end-to-end result and ablation study, to demonstrate the effectiveness of AnisoLift. The authors' contribution can be summarized as follows: 1. Formulated fixed-budget coarse-particle liquid enhancement. 2. Proposed AnisoLift and its aligning methods. 3. Validated AnisoLift on various benchmarks.

Critical strengths

The idea of enhance the represent ability of particle, rather than relying on additional particle to enhance simulation is valuable. Considering anisotropy rather than solely relying on weighted sum would be another major strength.

Critical weaknesses

It's unclear how the enhancement made by AnisoLift could be benefitial to downstream application. Though experimental results may have shown that, the mechanism remains ambiguous.

Questions for authors
  1. How does AnisoLift's enhancement help simulation? Section 3.1 states that 'residuals are applied only to the current observed frame and are not fed back into the solver', therefore it seems that the covariance ellipsoids can hardly refine the prediction result.
  2. Section 3.5 shown the final training objective as formulation 24. However, section 4.4 introduces a 'residual constrain', which is not included in former objective. Meanwhile, this 'residual constrain' seems to have exactly the same effect as state constrain. Please explain this conflict.
1 comment

We sincerely thank the reviewer for recognizing that enriching existing particles, rather than adding particles, is the point of the work, and for two questions that exposed a real gap in how we presented it. We answer them in two parts.

W1 and Q1. What does AnisoLift enhance, and how does it benefit downstream applications?

Scope and mechanism. AnisoLift does not make the coarse solver itself more accurate. Its output is not fed back, so the coarse trajectory is unchanged. What it changes is what each stored coarse frame represents. For every coarse particle it outputs (i) a corrected position and velocity (x^i,v^i)(\hat{x}_i, \hat{v}_i), closer to the high-resolution solution in the same region, and (ii) a covariance CiC_i describing how the high-resolution fluid around that particle is spread (extent and orientation), which an isotropic coarse particle cannot express. The benefit therefore goes to whatever consumes particle frames, not to the time integration: surface reconstruction and rendering, grid-field and free-surface analysis, and archived coarse datasets.

Downstream benefits. Most consumers of particle simulations, including rendering, surface reconstruction, and flow analysis, operate on the saved frames rather than inside the solver, so they directly benefit from enhanced states: (1) Physical analysis. Corrected velocities yield more faithful global quantities, e.g. kinetic-energy error reduced on 2D and 3D benchmarks (Table 1); (2) Efficiency. AnisoLift adds < 10 µs per frame, so coarse simulation plus AnisoLift remains 4.18× faster than the high-resolution solver (Table 8), and it can be applied selectively to frames of interest; (3) Surface reconstruction / rendering. The learned covariance C can be used directly as an anisotropic kernel for surface reconstruction (cf. Yu & Turk, 2013), without heuristic PCA-based estimation. According to the Table R2-1 and R2-2 below, our AnisoLift outperforms other baseline methods in IoU and F1 score, and this prove AnifoLift's excellence in Surface reconstruction / rendering.

Table R2-1. Occupancy reconstruction from particles and covariances on 2D DAM (mean ± std) across multiple random experiments. For each model, all rows except the first use that model's corrected centers, and all rows share one threshold.

Centers + covarianceIoU↑F1↑
Coarse + Yu & Turk0.735 ± 0.0040.847 ± 0.003
Corrected + Yu & Turk0.826 ± 0.0050.904 ± 0.003
Corrected + isotropic discs0.844 ± 0.0030.924 ± 0.002
Corrected + predicted ellipsoids (ours)0.873 ± 0.0040.932 ± 0.002

Table R2-2. Non-learned and learned baselines on the same targets on 2D DAM SPH simulation. MSEs follow Table 1. Learned models: mean ± std over 3 seeds. Laplacian smoothing (Yu & Turk) is applied to the coarse particles; each column takes the best of 18 kernel/radius/λ settings. Chamfer and IoU are measured against the high-resolution reference. Lower is better except IoU.

MethodMSE$_x $ ↓MSE$_v$ ↓MSE$_{Ekin}$ ↓Chamfer ↓Occupancy IoU ↑
Non-learned baseline
Coarse2.571e-33.567e-22.489e-40.04310.708
Laplacian smoothing (Yu & Turk)2.333e-33.308e-22.705e-40.04400.706
Learned baseline
Kernel regression (Nadaraya–Watson)1.122e-3 ±1.5e−5_{\pm 1.5e-5}1.560e-2 ±4.6e−5_{\pm 4.6e-5}2.761e-4 ±8.3e−6_{\pm 8.3e-6}0.0361 ±9.0e−4_{\pm 9.0e-4}0.780 ±4.3e−3_{\pm 4.3e-3}
Message-passing corrector (GNS-style)9.030e-4 ±2.2e−5_{\pm 2.2e-5}1.559e-2 ±4.9e−4_{\pm 4.9e-4}8.81e-5 ±2.1e−6_{\pm 2.1e-6}0.0357 ±8.7e−6_{\pm 8.7e-6}0.782 ±4.5e−3_{\pm 4.5e-3}
Ours, MLP 2×64 (submitted)1.104e-3 ±2.1e−5_{\pm 2.1e-5}2.255e-2 ±5.1e−4_{\pm 5.1e-4}1.249e-4$_{ \pm 4.8e-6}$0.0363 ±8.0e−4_{\pm 8.0e-4}0.768 ±4.0e−3_{\pm 4.0e-3}
Ours, MLP 4×2568.920e-4 ±2.5e−5_{\pm 2.5e-5}1.422e-2 ±3.7e−4_{\pm 3.7e-4}7.730e-5 ±2.1e−6_{\pm 2.1e-6}0.0347 ±7.0e−4_{\pm 7.0e-4}0.786 ±4.0e−3_{\pm 4.0e-3}

Q2. Clarification for the 'residual constrain'.

We thank the reviewer for the careful reading and apologize for the inconsistency. The residual-consistency loss was used in all experiments but was omitted from Eq. (24) and Algorithm 1. It is defined on the whitened residuals predicted in Eq. (13). The reviewer is correct that the two constraints share the same optimum, their difference lies in optimization. Lx\mathcal{L}_x and Lv\mathcal{L}_v operate in physical units, where position and velocity residuals differ substantially in scale and individual components have different variances, so the gradients are dominated by large-variance components. The whitened term gives each component a unit-scale contribution. Combining both balances absolute accuracy in physical units with uniform fitting across components and this is reflected in Table 2. We will add the residual-consistency term to Eq. (24) and clarify its role in Section 3.5 of the revision.

Reviewer_LUWvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AnisoLift post-processes coarse particle liquid simulations. For each coarse fluid particle, a two-layer MLP predicts current-frame position and velocity residuals and a covariance matrix. Supervision targets are kernel-weighted means and covariances of nearby high-resolution reference particles. Corrections are applied frame by frame and never fed back into the solver. Experiments on 2D and 3D Taylor-Green vortex (TGV), dam break, lid-driven cavity (LDC), and reverse Poiseuille flow (RPF), with SPH and MPS data, show lower errors than the uncorrected coarse trajectories.

Critical strengths
  1. A practical fixed-budget setting. Enhancing coarse simulations without adding particles or modifying the solver is a reasonable goal, and the decoupled design is simple, stable, and cheap at inference.
  2. Consistent gains over the coarse input across benchmarks and two solver families, supported by surface diagnostics on dam break (Table 4).
Critical weaknesses
  1. Limited novelty. The targets, a kernel-weighted neighbor mean for repositioning and a weighted covariance for the ellipsoid, follow the same formulation as Yu & Turk (2013, Eqs. 6 and 9-10), which also does not feed the repositioned particles back into the simulation. The main difference is that the statistics are computed on reference particles and regressed by an MLP. Moreover, the predicted covariance is not used in the dynamics or in any quantitatively evaluated reconstruction, so its value beyond visualization is unclear.
  2. Weak evaluation. The reported errors are the training losses computed on the same constructed targets (Eqs. 18-19 and 27-28). The only baseline is the uncorrected coarse state; there is no non-learned baseline, such as Yu & Turk smoothing applied to the coarse particles, and no learned per-frame corrector trained on the same targets. LDC and RPF each use a single trajectory with a temporal split and absolute-position inputs, yet they yield the largest gains (about 300-fold on 2D LDC), which may reflect memorization. The problems are small (at most 8000 particles), some gains are negligible (3D TGV SPH position error from 0.0519 to 0.0517), and no variances are reported.

Minor suggestions for improvement (not affecting the overall score):

  1. The residual-consistency loss ablated in Table 2 does not appear in Eq. 24, and the isotropic variant is described differently in Section 4.4 and the Table 2 caption.
  2. The height error in Table 4 (0.707 to 0.112) does not match the per-frame values in Figure 5 (about 0.06 to 0.03).
  3. Table 8 excludes feature extraction, and the 4.18-fold speedup reflects only the solver resolution. Table 10 includes unexplained AnisoLift rows with 2800 and 4000 particles. There are also typos, such as "Appdix" and "hyperprameters".
Questions for authors
  1. Beyond using reference particles, how does the target construction differ from Yu & Turk (2013)? Please report a non-learned baseline that applies it directly to the coarse particles.
  2. Is the predicted covariance used in any quantitative evaluation, such as Table 4? If not, what shows that the ellipsoid itself, rather than an auxiliary-loss effect, improves the results?
  3. On LDC and RPF, how does performance change without absolute-position inputs or on a different trajectory?
6 comments

W1.1 and Q1.1. Clarification of the novelty and the difference between ours and Yu & Turk

We sincerely thank you for this concern. We agree the formulas are the same (their Eqs. 6 and 9–10), and we will credit this work in Section 3.3 (highlighted in blue in the revised PDF). The difference is the input. Yu & Turk smooth a particle using its own coarse neighbors, so on a coarse simulation their statistics describe only the coarse arrangement. Ours are computed from a high-resolution reference, which is unavailable at test time, and must be predicted from the coarse state. Our contribution is therefore the supervision: per-particle targets derived from a high-resolution reference on a fixed particle budget, and a corrector that predicts them from the coarse state. The results in W2.2 and Q1.2 confirm that the improvement comes from this difference in input, not from the form of the statistics.

W1.2. The use and value of covariance.

Thank you for caring about the usage and the value of the covariance in our method. Keeping the covariance out of the dynamics is is a design choice, AnisoLift post-processes a precomputed coarse trajectory without feeding back into the solver, which keeps it solver-agnostic and stable, as the reviewer notes among the strengths. The covariance's value lies in what the corrected state can represent. With it, a coarse particle stands for an anisotropic region of the high-resolution flow rather than a point (Section 3.4), and the pair (corrected center, covariance) is exactly the input of anisotropic-kernel surface reconstruction, the use Yu & Turk designed their covariance for. We now evaluate it quantitatively in that setting (Table R1-1). With the centers fixed, predicted ellipsoids raise occupancy IoU over isotropic discs (0.844 → 0.873), while Yu & Turk covariances computed from the coarse particles lower it (0.797). So better covariance prediction translates directly into better reconstruction. Q2 gives the full comparison. We highlight the corresponding manuscript change in blue in the second paragraph in the section 4.4 of the revision.

Q2. Is the predicted covariance used in any quantitative evaluation? What shows that the ellipsoid, rather than the auxiliary loss, drives the improvement?

Thank you for questions about the exact improvement mechanism of our method. The two outputs of AnisoLift have separate roles (Section 3.4), the corrected centers and velocities carry the dynamics, and the covariance carries the local geometry of the reference flow, which a point sample cannot represent. The ellipsoid's contribution should therefore be measured where this geometry is used, in reconstructing the fluid region, and in a way that no training loss can account for. In Table R1-1 (also appears in the response to the reviewer Gy7z, and marked as Table R2-1), predicted ellipsoids reach higher IoU and F1 score compared with other variants (isotropic discs and Yu & Turk covariances). The centers are held fixed and only the covariance changes between rows, which isolates the ellipsoid from every training loss. We highlight the corresponding manuscript change in blue in the second paragraph in the section 4.4 of the revision.

Table R1-1. Occupancy reconstruction from particles and covariances on 2D DAM (mean ± std) across multiple random experiments. For each model, all rows except the first use that model's corrected centers, and all rows share one threshold.

Centers + covarianceIoU↑F1↑
Coarse + Yu & Turk0.735 ± 0.0040.847 ± 0.003
Corrected + Yu & Turk0.826 ± 0.0050.904 ± 0.003
Corrected + isotropic discs0.844 ± 0.0030.924 ± 0.002
Corrected + predicted ellipsoids (ours)0.873 ± 0.0040.932 ± 0.002

Moreover, predicted ellipsoids beat isotropic discs on all test trajectories for every learned model in Table R1-2.

W2.1. Reported errors are the training losses computed on the same constructed targets.

Thank you for your comments on the evaluation protocol of our method. Similar to learned Lagrangian simulators such as GNS and Neural SPH, our method is supervised using high-resolution simulation data.. However, unlike these methods, our coarse and high-resolution particles do not have one-to-one correspondences. We therefore construct a supervision target for each coarse particle by projecting the high-resolution reference onto it through a kernel-weighted average of its neighboring high-resolution particles, following the particle-transfer approach of Yu and Turk (2013). Table 1 measures how well this projection is recovered. Notably, our training objective already incorporates a reference-based constraint through the kinetic-energy term (Section 3), which directly compares the kinetic energy of the corrected states against that of the high-resolution reference, rather than against the projected supervision targets.

W2.2 and Q1.2. Extended comparison with unlearned and learned baselines

Thank you for your suggestions about a wide range of evaluation experiments on various baselines. We trained two standard regressors on exactly our targets and inputs: Nadaraya–Watson kernel regression (k and bandwidth chosen on validation) and a one-round message-passing corrector with a GNS-style encoder and our head and loss. We also trained a wider per-particle MLP (4×256, chosen on validation) for AnisoLift. According to the Table R1-2 (also appears in the response to reviewer Gy7z and marked as Table R2-1), on 2D DAM, every learned model improves on the coarse state against the reference as well, in Chamfer distance and IoU (Note that this IoU differs from the free-surface IoU in Table R1-1. Here, each particle covers a fixed radius. There, each particle contributes a kernel shaped by its covariance, and the kernel scale is chosen once per trajectory to maximize IoU and shared by all rows. The values are therefore lower, and comparable only within this table.) on all test trajectories, whereas Laplacian smoothing does not. The stronger backbones (ours, MLP 4$\times$256) improve further on the submitted 2×64 MLP (Table R1-2): the message-passing corrector and the 4×256 MLP reach MSE$_x$ 9.030e-4 and 8.920e-4 vs. 1.104e-3, MSE$_v$ 1.559e-2 and 1.422e-2 vs. 2.255e-2, and kinetic-energy error 8.81e-5 and 7.730e-5 vs. 1.249e-4 (three seeds; every difference exceeds one standard deviation). Kernel regression, which has no kinetic-energy term, lowers MSE$_v$ but leaves the kinetic-energy error above that of the coarse state, highlighting the benefit of explicitly incorporating the kinetic-energy term in our objective. Taken together, these results support that our contribution lies in the reference-derived, per-particle supervision targets under a fixed coarse-particle budget. All learned regressors benefit from these targets, while stronger architectures can exploit them further. We highlight the corresponding manuscript change in blue in the baselines and target-independent metrics paragraph in section 4.2 of the revision PDF.

Table R1-2. Non-learned and learned baselines on the same targets on 2D DAM SPH simulation. MSEs follow Table 1. Learned models: mean ± std over 3 seeds. Laplacian smoothing (Yu & Turk) is applied to the coarse particles; each column takes the best of 18 kernel/radius/λ settings. Chamfer and IoU are measured against the high-resolution reference. Lower is better except IoU.

MethodMSE$_x $ ↓MSE$_v$ ↓MSE$_{Ekin}$ ↓Chamfer ↓Occupancy IoU ↑
Non-learned baseline
Coarse2.571e-33.567e-22.489e-40.04310.708
Laplacian smoothing (Yu & Turk)2.333e-33.308e-22.705e-40.04400.706
Learned baseline
Kernel regression (Nadaraya–Watson)1.122e-3 ±1.5e−5_{\pm 1.5e-5}1.560e-2 ±4.6e−5_{\pm 4.6e-5}2.761e-4 ±8.3e−6_{\pm 8.3e-6}0.0361 ±9.0e−4_{\pm 9.0e-4}0.780 ±4.3e−3_{\pm 4.3e-3}
Message-passing corrector (GNS-style)9.030e-4 ±2.2e−5_{\pm 2.2e-5}1.559e-2 ±4.9e−4_{\pm 4.9e-4}8.81e-5 ±2.1e−6_{\pm 2.1e-6}0.0357 ±8.7e−6_{\pm 8.7e-6}0.782 ±4.5e−3_{\pm 4.5e-3}
Ours, MLP 2×64 (submitted)1.104e-3 ±2.1e−5_{\pm 2.1e-5}2.255e-2 ±5.1e−4_{\pm 5.1e-4}1.249e-4$_{ \pm 4.8e-6}$0.0363 ±8.0e−4_{\pm 8.0e-4}0.768 ±4.0e−3_{\pm 4.0e-3}
Ours, MLP 4×2568.920e-4 ±2.5e−5_{\pm 2.5e-5}1.422e-2 ±3.7e−4_{\pm 3.7e-4}7.730e-5 ±2.1e−6_{\pm 2.1e-6}0.0347 ±7.0e−4_{\pm 7.0e-4}0.786 ±4.0e−3_{\pm 4.0e-3}

W2.3 and Q3. Memorization check on 2D LDC and 2D RPF and how does performance change on LDC and RPF without absolute-position inputs, or on a different trajectory?

Thank you for your questions about the experiments on 2D LDC and 2D RPF scenes. On 2D LDC, the gain is not memorization of the training trajectory: it transfers to a new Reynolds number, and a clear correction remains without any positional input. We regenerated both scenes and tested both changes (Table R1-3 a and b). With absolute positions, the position gain is 69 ± 18× on an independent rollout of the training configuration and 54 ± 16× on a new trajectory at Re = 200 (training at Re = 100). Without absolute positions, it falls to 2.7× and 2.4×, while velocity still improves about 2× and kinetic energy 5.9× and 11.1×. This removes all positional information, including wall distances that a translation-invariant model would legitimately use, so it is a conservative test. The size of the LDC factor has a simple cause. The coarse position error is almost the same at Re = 100 and Re = 200 (5.43e-4 and 5.44e-4), so it is dominated by a systematic, geometry-induced offset rather than by flow details; a position-aware corrector removes most of this offset, which is why the factor is large, drops without positions, and still transfers across Reynolds number. Table 1 reported a roughly 320×, on the regenerated data averaged over three seeds, the factor is 69 ± 18×, and our revision will report this value in the updated Table 1. This factor thus reflects a geometry-specific correction, and the revision will present LDC as a within-geometry result. On 2D RPF, a similar situation exits (shown in Table R1-3 b), and the 100$\times$ enhancement is also due to trivial difference between coarse and reference, and the revision will present that 2D RPF enjoys 3.7$\times$, 1.34$\times$, and 1.06$\times$ improvement on position, velocity and kinetic energy.

Table R1-3. Memorization check on 2D LDC and 2D RPF. Each model is trained on a single trajectory with three seeds. MSEs follow Table 1 and are mean ± std over seeds; gain = coarse error / corrected error, computed per seed, then mean ± std (> 1 means improvement). Coarse rows are solver outputs and have no seed variance. "Without abs. position" uses inputs [v, ρ, p].

(a) 2D LDC, trained at Re = 100

Test trajectoryModelMSE$_x$↓Position gainMSE$_v$↓Velocity gainMSE$_{Ekin}$↓Kinetic-energy gain
Independent rollout, Re = 100Coarse5.43e-4—2.35e-2—1.24e-5—
Independent rollout, Re = 100With abs. position8.28e-6 ± 2.6e-669 ± 18×4.01e-3 ± 2.1e-45.9 ± 0.3×2.60e-6 ± 3.5e-74.9 ± 0.7×
Independent rollout, Re = 100Without abs. position2.00e-4 ± 4.4e-72.712 ± 0.006×1.07e-2 ± 2.5e-52.186 ± 0.005×2.12e-6 ± 1.4e-75.9 ± 0.4×
New trajectory, Re = 200Coarse5.44e-4—2.43e-2—3.36e-5—
New trajectory, Re = 200With abs. position1.09e-5 ± 4.0e-654 ± 16×4.47e-3 ± 1.8e-45.5 ± 0.2×5.34e-6 ± 6.8e-76.4 ± 0.8×
New trajectory, Re = 200Without abs. position2.23e-4 ± 3.2e-72.440 ± 0.004×1.27e-2 ± 8.1e-51.91 ± 0.01×3.03e-6 ± 1.7e-711.1 ± 0.6×

(b) 2D RPF, trained with body force 1

Test trajectoryModelMSE$_x$Position gainMSE$_v$Velocity gainMSE$_{Ekin}$Kinetic-energy gain
Independent, force 1Coarse7.80e-5—4.75e-3—4.24e-5—
Independent, force 1With abs. position2.11e-5 ± 2.9e-63.70±0.05×3.54e-3 ± 7.6e-41.34±0.09×3.99e-5 ± 1.2e-51.06±0.01×
Independent, force 1Without abs. position4.15e-5 ± 1.4e-71.88±0.06×3.72e-3 ± 1.3e-31.28±0.004×4.12e-5 ± 1.0e-51.03±0.002×
New trajectory, force 2Coarse7.81e-5—4.58e-2—5.32e-4—
New trajectory, force 2With abs. position6.60e-5 ± 1.6e-51.18±0.03×3.64e-2 ± 7.2e-31.26±0.02×4.58e-4 ± 5.3e-51.16±0.007×
New trajectory, force 2Without abs. position7.06e-5 ± 4.0e-71.11±0.002×4.36e-2 ± 7.1e-31.05±0.001×5.14e-4 ± 6.4e-51.04±0.003×

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary for the Area Chair and Reviewers

We thank the Area Chair and reviewers for their constructive feedback. In response, we have substantially strengthened the experimental evidence supporting AnisoLift's core contributions and incorporated the requested analyses and clarifications into the revised manuscript.

Core contribution. Given a coarse particle simulation with a fixed particle budget, AnisoLift predicts per-particle position and velocity corrections together with an anisotropic shape representation, operating as a post-processing module decoupled from the underlying solver. Its key contribution is to jointly learn the corrected particle states and their anisotropic shapes, enabling each coarse particle to represent a larger and more faithful region of the fine-scale liquid geometry at negligible computational overhead.

Main clarifications and supporting evidence:

  1. Evaluation independent of our targets. On 2D dam break scene, our corrected state is closer to the high-resolution particles on every metric, including Chamfer distance, free-surface IoU, surface-height MAE, and kinetic-energy error.

  2. Extended baselines. We extend comparison to non-learned (Laplacian smoothing in Yu & Turk) and learned (Kernel regression, GNS corrector) baselines, and our strengthened AnisoLift improve all metrics further than others.

  3. Contribution of the anisotropic shape. We reconstruct liquid occupancy from the same corrected centers, our predicted ellipsoids reach better surface reconstruction performance than isotropic discs and covariances in Yu & Turk.

  4. Temporal behaviour and conservation. We evaluate AnisoLift's conservation properties according to the underlying physics of simulation scenes. The results show that AnisoLift's corrected trajectories enjoy lower relative error of momentum and kinetic energy than coarse trajectories.

  5. Computation cost. Given a coarse particle simulation with a fixed particle budget, AnisoLift predicts per-particle position and velocity corrections together with an anisotropic shape representation, operating as a post-processing module decoupled from the underlying solver. Its key contribution is to jointly learn the corrected particle states and their anisotropic shapes, enabling each coarse particle to represent a larger and more faithful region of the fine-scale liquid geometry at negligible computational overhead.

  6. Possible memorization. We evaluate AnisoLift by inputting without absolute positions on 2D LDC and 2D RPF, and testing on target-independent metrics and on different domains, reliable enhancement results are still observed over multiple random experiments.

The revised main manuscript presents the corresponding findings and conclusions, while detailed experimental protocols and results are provided in the appendices. All modifications to the manuscript are highlighted in blue, and each response identifies the corresponding revised sections, tables, and paragraphs, where applicable.

Dear Reviewers,

Thank you for taking the time to review our paper. As the discussion period draws to a close, we’d appreciate any feedback on whether our response addressed your concerns. We’d be happy to clarify any remaining points.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.