acceptodds
Under review as a conference paper at ICLR 2027

Conservative Local-Global Transport: Mesh-Local Convolution and Scalable Linear Attention for Neural Lagrangian Simulation

Abstract

Many neural Lagrangian simulators rely on Graph Neural Networks (GNNs) with spatially constructed interaction graphs, introducing considerable topology-processing and message-passing costs as the particle count grows. Moreover, these architectures are typically agnostic to the underlying conservation laws. In this work, we propose a conservative local–global transport mechanism in latent space: mesh convolution captures strongly coupled near-field interactions, while separable linear attention enables dense long-range interactions at linear complexity, both enforcing conservation laws by construction. Beyond global conservation, we introduce sign-split latent flux limiters that strictly preserve prescribed physical bounds without breaking the conservative latent update, and connect these discrete limiters to a unified family of latent advection dynamics. It demonstrates advantages over GNN-based baselines in computational efficiency (supports up to more particles) and physical consistency while maintaining competitive predictive accuracy. Our code is available at https://anonymous.4open.science/r/CONSERVATIVE-LOCAL-GLOBAL-TRANSPORT-I16S.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural Lagrangian simulation, structure-preserving neural solver, scalability, latent conservation, linear attention

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rTbROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper makes linear attention skew-symmetric, so pairwise latent fluxes cancel and any linearly projected quantity is conserved exactly at O(N) cost. It adds a conservative grid-based MeshConv for near-field interactions and a sign-split flux limiter that enforces bounds such as non-negative concentration. On 2D TGV and lid-driven cavity, the method conserves mass and momentum almost exactly and scales to about 38× more particles than graph-based baselines, at somewhat lower accuracy than GNS.

Critical strengths

The core idea is simple and meaningful: exact conservation by construction, linear complexity, and no graph construction. The limiter preserves both skew-symmetry and O(N) evaluation. The scalability gains are clear and practically useful.

Critical weaknesses

Closely related baselines are missing: Horie & Mitsume (2024), which uses antisymmetric messages for conservation, and Transolver (Wu et al., 2024), a linear-complexity transformer for PDEs. Both are cited but not compared. The evaluation is limited to two small 2D cases with 2,500 particles. The 3D result is qualitative only, and accuracy is below GNS. Quantitative results on a standard benchmark such as LagrangeBench, or at least 3D error against ground truth, would strengthen the paper considerably.

Questions for authors

Is the limiter used during training or only at inference? Also, which Table 2 configuration matches the full model in Table 1? The concentration losses don’t line up.

Reviewer_5qNaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a conservative local–global transport framework for neural Lagrangian simulation. It combines skew-symmetric linear attention for long-range interactions with a conservative mesh convolution for local interactions, and introduces sign-split latent flux limiters to enforce prescribed physical bounds while preserving conservation. Experiments demonstrate improved physical consistency.

Critical strengths
  • The proposed method incorporates conservation constraints directly into the architecture.
  • The proposed method outperforms baselines in some cases.
Critical weaknesses
  1. The presentation of the method is difficult to follow in several places. For example, Eqs. 7–9 use x+x^+, x−x^-, L+L^+, L−L^-, R+R^+, and R−R^-, but the paper does not explicitly define the positive and negative parts or explain what the ++ and −- stand for. Some other notation is also introduced implicitly: τi=Δt/Vi\tau_i=\Delta t/V_i is used in the limiter without much motivation, and pkp_k in Appendix B could be defined more explicitly. By the way, is pkp_k explicitly linked to any physical quantity like momentum? Since pp is reused to define the momentum at L194. These presentation issues make the derivations harder to follow and verify.

  2. Section 3.1 states that spectral constraints bound the Lipschitz constant of gθg_\theta by γ<1\gamma<1, but the paper does not provide a derivation for the full update map or explain how the bound depends on the attention normalization, time step, particle volumes, and source mapping. The authors should provide the bound and clarify how it is enforced and verified in the implementation.

  3. In the main text, there is no explicit explanation of how the concentration cc is updated. According to the "Limitation" right before Section 3.2, the limiter should be the one that address the conservation of the concentration. However, in the appendix Equation 87, the concentration is simply updated by a projection. What's the relation between the Equation 87 in Appendix and the Section 3.2? How the Equation 87 address the conservation? In addition, the proof in Appendix A.2 is for a source-free update. Since the reactive setting includes reaction sources, the authors should clarify how those sources are handled to preserve nonnegative species concentrations.

  4. The reported results suggest an accuracy–efficiency tradeoff that deserves more discussion. On 2D-TGV, the proposed method’s fluid and concentration losses are approximately 3.7×3.7\times and 4.6×4.6\times those of GNS, respectively. On 2D-LDC, the corresponding losses are also higher than GNS. These results should be reconciled with the claim of “competitive accuracy” at L73. The scalability experiments demonstrate memory/runtime behavior at larger particle counts, but do not establish predictive accuracy at those scales. Evaluating accuracy or rollout stability as particle count increases would strengthen the scalability claim.

  5. The supplementary visualizations appear to focus on results at different resolutions. Side-by-side rollout visualizations for the baselines would help assess temporal behavior and visual quality, and a 3D dam-break rollout would be particularly useful. In the selected 3D frames, the prediction appears smoother than the ground truth; a full rollout would help determine whether this reflects temporal oversmoothing or simply the selected frames.

Questions for authors

Please refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Main Reply

Q1 Notation and flux scaling

A: Thank you for pointing out the missing definitions. First, for a scalar xx, we define x+=max⁡(x,0)x^+=\max(x,0) and x−=max⁡(−x,0)x^-=\max(-x,0). Both parts are nonnegative, and x=x+−x−x=x^+-x^-. The same convention applies entrywise to LL and RR.

Second, τi=Δt/Vi\tau_i=\Delta t/V_i, where Δt\Delta t is the transport time step and Vi=mi/ρ0V_i=m_i/\rho_0 is the reference volume of particle ii. Dividing both sides of Eq. (13) by ViV_i gives this factor, which converts a flux into a change in the particle state over one step.

Third, in Appendix B.1, pkp_k denotes the locally frozen scalar attention coefficient for the interaction from grid point i−ki-k to grid point ii, where kk is a grid offset. At L194, pi=WFzFi\mathbf p_i=W_F\mathbf z_{Fi} denotes the momentum vector of particle ii.

Q2 Contraction bound

A: The reported experiments use a fixed number of Picard-style latent updates with the hard contraction condition left unenforced, since Hard contraction constraints can limit the expressiveness of the learned transport operator (Baker et al., ICML 2023) which is also verified in our early attempts. Therefore, the Picard updates is downgraded to a design intention and inductive bias in a pipeline level. Its effectiveness is verified by Fig. 5(b) where iterations with proper times help to improve the training dynamics. We have clarified this in the updated draft. Furthermore, such update rule is optional since this has nothing to do with our two main claims: conservation and scalability.

Q3 Conservation and nonnegativity

A: We summarize our response into two points as it is a bit lengthy. We will cover these implementation details in our work.

  • For the transport, the limiter in Sec. 3.2 scales the latent fluxes while preserving their pairwise antisymmetry. The shared linear decoder in Eq. (87) maps the accumulated limited increment to the concentration increment, preserving the total mass of each species during transport. The proof you refer to is for Theorem A.2 in Appendix A.1 but not Appendix A.2. It assumes a source-free gated transport update.
  • For the source term in chemical reactions, it is different from momemtum source since it still follows atomic conservation governed by the stoichiometric matrix S\mathbf S while the other does not. After transport, we cap the total consumption of each species at the amount available and update the species carried by a particle in a ratio with accordance to S\mathbf S. By limiting the reaction rate at each time step, the non-negativity is preserved.

Q4 Accuracy and scalability

A: GNS achieves lower fluid and concentration prediction losses on both datasets. Our model runs approximately 3.2×3.2\times as fast as GNS on 2D-TGV and 3.6×3.6\times as fast on 2D-LDC. It also achieves lower fluid and concentration prediction losses than SFBC and FluidFormer on both datasets and has the lowest runtime in Table 1. We propose describing these results as a favorable accuracy–efficiency trade-off in the revision.

The experiments in Fig. 3b assess hardware scalability through runtime and peak memory. We evaluate rollout consistency under denser particle sampling against a reference interpolated from the coarse trajectory. For 2D TGV, we use a split-and-interpolate scheme to increase the particle count from 2,500 to 1.715M and run the model trained at the original particle count without retraining.

The table below reports position MSE ( Lfluid\mathcal L_{\text{fluid}} ) and concentration MSE (Lconc\mathcal L_{\text{conc}}). The revised manuscript also includes the corresponding multi-resolution rollout visualizations in Appendix D (Fig. 6).

ParticleCountPositionMSE(Lfluid\mathcal{L}_{\text{fluid}})ConcentrationMSE(Lconc\mathcal{L}_{\text{conc}})Runtime(ms/step)
2,5007.192×10−47.192\times10^{-4}6.220×10−56.220\times10^{-5}11.98
1.715M2.951×10−32.951\times10^{-3}1.202×10−41.202\times10^{-4}6,055.85
  • Remark: We measure both runtimes on the same NVIDIA RTX A6000 GPU.

All baselines run out of GPU memory at 1.715M particles, which precludes evaluation at this scale.

Q5 Visualization

A: The full TGV rollout animation appears in the README of the anonymous code repository which also includes side-by-side rollout GIFs comparing our model with the baselines on 2D-TGV and 2D-LDC, along with the full 3D dam-break rollout. Visual inspection of the 3D rollout suggests mild spatial smoothing whose severity appears broadly similar over time. Deterministic MSE training may contribute to this effect, particularly in isolated splashes and fine-scale particle structure.

AuthorsOpenReview ↗

Details of Reply for Q2,Q4

Q2

Below, we derive a Lipschitz bound and show how an optional implementation can satisfy the contraction condition in our previous attemp. The conservation and separable aggregation properties follow from the transport construction itself.

For a state Z=(zi)i=1N\mathbf Z=(\mathbf z_i)_{i=1}^N, let Mi(Z)M_i(\mathbf Z) and Ai(Z)A_i(\mathbf Z) denote the MeshConv and ConsAttn flux terms inside the brackets in Eq. (13), respectively. Let Φi(Z;fi)\Phi_i(\mathbf Z;f_i) denote the source rate at particle ii for the prescribed source input fif_i. With the transport sign convention of Eq. (13), the Picard update is

Ui(Z)=zi(0)+ΔtVi[−(Mi(Z)+Ai(Z))+Φi(Z;fi)].U_i(\mathbf Z)=\mathbf z_i^{(0)}+\frac{\Delta t}{V_i}\left[-\bigl(M_i(\mathbf Z)+A_i(\mathbf Z)\bigr)+\Phi_i(\mathbf Z;f_i)\right].

The Lipschitz constant of the full update satisfies Lip⁡(U)≤max⁡i(ΔtVi)(LM+LA+Lf).\operatorname{Lip}(U)\le \max_i\left(\frac{\Delta t}{V_i}\right)(L_M+L_A+L_f).

ConsAttn. Let qiq_i,kik_i,viv_i be the mapped query, key, and value features of particle ii. Denote their norm bounds by RqR_q,RkR_k,RvR_v and their Lipschitz bounds by LqL_q,LkL_k,LvL_v. These bounds can be imposed through bounded normalization, spectral clipping of the projections, and bounds on the kernel feature map. The pairwise score sij=qi⊤kjs_{ij}=q_i^\top k_j has magnitude at most B=RqRkB=R_qR_k and Lipschitz constant at most Ls=RkLq+RqLkL_s=R_kL_q+R_qL_k. With a numerical stabilizer ϵ>0\epsilon>0, one can impose a denominator floor of NδN\delta for δ>0\delta>0 by defining the normalized score as

aij=sijDi,Di=max⁡(∑j=1Nsij+ϵ, Nδ).a_{ij}=\frac{s_{ij}}{D_i}, \qquad D_i=\max(\sum_{j=1}^{N}s_{ij}+\epsilon,\ N\delta).

The denominator has Lipschitz bound NLsNL_s. Taking the difference of two quotients and using Di≥NδD_i\ge N\delta, we obtain

Lip(aij)≤LsN(1δ+Bδ2) and ∣aij∣≤BNδ.\text{Lip}(a_{ij}) \le \frac{L_s}{N}\left(\frac{1}{\delta}+\frac{B}{\delta^2}\right)\text{ and } |a_{ij}|\le \frac{B}{N\delta}.

Let gAg_A be the attention gate. Applying the product bound to each directed term in Ai=gA∑j(aijvj−ajivi)A_i=g_A\sum_j(a_{ij}v_j-a_{ji}v_i) gives

LA≤2∣gA∣[BδLv+RvLs(1δ+Bδ2)].L_A\le 2|g_A|\left[ \frac{B}{\delta}L_v+ R_vL_s\left(\frac{1}{\delta}+\frac{B}{\delta^2}\right) \right].

The factor of two accounts for the two directions of the conservative flux. The factor 1/N1/N cancels when summing over particles. MeshConv. Let PP be the nonnegative scatter matrix and P⊤P^\top the corresponding gather operator. Each column of PP sums to at most one. One can cap each row sum at ρ∗\rho_* by rescaling the scatter weights at grid nodes where the row sum exceeds ρ∗\rho_*. Let Kc[ξ]K_c[\xi] be the effective convolution coefficient for channel cc at offset ξ\xi. One can also impose the kernel bound ∑ξmax⁡c∣Kc[ξ]∣≤κ∗\sum_\xi\max_c|K_c[\xi]|\le\kappa_* for a prescribed κ∗\kappa_*. Under these bounds, the forward grid transfer and the reverse-flux correction each have operator norm at most ρ∗κ∗\rho_*\kappa_*. The normalization and the input and output projections can each be constrained to be 1-Lipschitz. Under these conditions, the MeshConv gate gMg_M gives

LM≤2∣gM∣ρ∗κ∗.L_M\le 2|g_M|\rho_*\kappa_*.

Source term. If Φi\Phi_i depends only on the fixed input fif_i, its Lipschitz contribution is Lf=0L_f=0. One possible state-dependent source map is Φi(Z;fi)=Woσ(Wzzi+Wffi)\Phi_i(\mathbf Z;f_i)=W_o\sigma(W_z\mathbf z_i+W_f f_i), where WzW_z and WfW_f project the latent state and source input, WoW_o is the output projection, and σ\sigma is an activation with Lipschitz bound LσL_\sigma. With fif_i held fixed, we obtain

Lf≤Lσ∥∥Wz∥∥2∥∥Wo∥∥2.L_f \le L_{\sigma} {\|\| W_z\|\|}_2 {\|\|W_o\|\|}_2.

For a matrix W∈Rm×nW\in\mathbb R^{m\times n}, the spectral norm ∥∥W∥∥2\|\|W\|\|_2 is its largest singular value.

Example of a contractive implementation. For a source-free update, Lf=0L_f=0. Periodic 2D-TGV is an example of a source-free setting. With equal reference volumes Vi=VV_i=V, one can absorb the common factor Δt/V\Delta t/V into the transport outputs. As an example, take unit feature-norm bounds, (Lq,Lk,Lv)=(0.25,0.25,1)(L_q,L_k,L_v)=(0.25,0.25,1), and δ=0.5\delta=0.5, giving LA≤10∣gA∣L_A\le 10|g_A|. Setting ρ∗=2\rho_*=2 and κ∗=1\kappa_*=1 gives LM≤4∣gM∣L_M\le 4|g_M|. One can then cap the gates at ∣gA∣≤0.04|g_A|\le 0.04 and ∣gM∣≤0.10|g_M|\le 0.10 to obtain the theoretical upper bound

Lip⁡(U)≤10∣gA∣+4∣gM∣≤0.4+0.4=0.8<1.\operatorname{Lip}(U) \le 10|g_A|+4|g_M| \le 0.4+0.4=0.8<1.

Q4

To construct the finer representation, we split each of the 2,500 coarse particles into 686 children via stratified random sampling. Each child receives 1/6861/686 of its parent particle's mass. We conduct SPH relaxation on the first frame, then advance the child states over the initial six frames using fixed-weight quintic-kernel interpolation from the coarse trajectory. During the subsequent rollout, the gate coefficient of MeshConv is scaled by 1686\frac{1}{686} to compensate for its sum-based scatter aggregation. For the denser particle set, we compute quintic-kernel interpolation weights for each child particle after the initial SPH relaxation. At each subsequent frame, we use these fixed weights to construct the reference positions from cumulative coarse-particle displacements and to interpolate the reference concentrations.

Reviewer_4SWQOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

A neural Lagrangian simulator is built from two conservative latent operators, a skew-symmetrized linear attention and a particle-grid convolution, plus an O(N)O(N) sign-split flux limiter that enforces bounds. On self-generated 2D SPH flows (TGV and LDC) it reports near-zero mass error, zero negativity, lowest runtime, and about 38x more particles in the same GPU memory than GNS, EGNN, SFBC and FluidFormer, at higher error than GNS.

Critical strengths
  • The guarantees (i.e. conservation and bound preservation) are correct and hold in practice.
  • Cost scales well.
Critical weaknesses
  • Many claims are unsupported/overstated.
  • The theory apart from the two proofs provided does not match the trained model.

See questions for further details.

Questions for authors

Presentation

Clarifications

  1. Which model do Tables 2a-c ablate? Their shared base (Lconc=3.406×10−3\mathcal{L}_{\text{conc}} = 3.406\times10^{-3} in 2a "None", 2b "None" and 2c 64×6464\times64) appears to lack both spatial bias and limiter, while the Table 1 model (0.064×10−30.064\times10^{-3} on TGV) is closest to 2b "C" (0.057×10−30.057\times10^{-3}). There, the limiter cuts Lconc\mathcal{L}_{\text{conc}} ~4x, despite "rather than to increase predictive capacity" (Sec. 4.3); what does it do for the reported model?
  2. Are Lfluid\mathcal{L}_{\text{fluid}} and Lconc\mathcal{L}_{\text{conc}} single-step or rollout errors, and at what horizon are EmomE_{\text{mom}}, EmassE_{\text{mass}}, M−M_- and r−r_- computed?
  3. What does the main text's 3D count, "about 60K", include, given App. C.2's 52,488 fluid particles?

Errata

  1. Table 2's caption gives units for an EmassE_{\text{mass}} column that Table 2a lacks, though Sec. 4.3 cites "nearly identical prediction and conservation errors", and its r−r_- unit of 10010^{0} makes the "None" row a fraction of 4.917. Where are the conservation errors, and is the unit 10−210^{-2} as in Table 1?
  2. The LagrangeBench reference lists "Makoto Takamoto, Krishnan Pradip, et al." (Takamoto is PDEBench's first author) and a title matching neither paper; LagrangeBench is by Toshev et al. (NeurIPS 2023). Which suite does Sec. 4.1 follow?
  3. Should Eq. 2's boundary term be written for a general flux rather than only the diffusive ∇u~⋅n\nabla\tilde{u}\cdot n?
  4. Should "Lid-Driven Capacity Flow" (App. C.2, C.3) read "Cavity"?

Soundness

Clarifications

  1. What exactly is conserved? Thm. A.1 proves an unweighted sum and Eq. 13 states a volume-weighted one, but neither the inputs (App. C.3) nor the decoder update (Eq. 87) involve density or volume, while the data are weakly compressible SPH (Vi=mi/ρiV_i = m_i/\rho_i). If conservation is exact, what produces Emass=1.0×10−4E_{\text{mass}} = 1.0\times10^{-4} on LDC, 20x TGV's and above typical roundoff?
  2. Was Eq. 25's latent stoichiometric constraint used in the 3D runs? A source c˙i=Sfir\dot c_i = S f^r_i "automatically preserves the atomic stoichiometry" in physical space, but Sec. 3.1 couples it in latent space.
  3. In Fig. 5b, was one model iterated longer or a model retrained per count? What LL was used in training, and what spectral constraint is implemented? Is the scheme a Picard iteration of Eq. 5 or LL explicit sub-steps?
  4. LagrangeBench, which Sec. 4.1 appears to follow (Q5), has TGV-2D (also 2,500 particles) and LDC-2D with published GNS and SEGNN results; it has no solute, but the fluid results could've been built on it. Why not use these datasets?

Errata

  1. Is "competitive predictive accuracy" accurate? The model is second of five in three of four loss columns, but has 3.7x and 4.6x GNS's error on TGV (fluid 0.775 vs 0.211; concentration 0.064 vs 0.014), 8% more on LDC fluid, and ranks third on LDC concentration (1.842 vs 1.040 and 1.150).
  2. EmomE_{\text{mom}} divides by ∥P∥\|P\|, which Sec. 4.1 does not define. If PP is total momentum, it is near zero on both benchmarks (by symmetry on periodic TGV; by the divergence theorem in a closed cavity with a tangential lid), so neither column is a meaningful relative error, as the TGV baselines' 44 to 4,671 show; on LDC the model and GNS tie (0.126 vs 0.130). What is PP, what is the absolute drift, and does the advantage that "remains in 2D-LDC" rest on anything beyond mass and negativity?
  3. Sec. 3.1 says spectral constraints bound the Lipschitz constant by γ<1\gamma < 1, "ensuring exponential convergence", without a derivation for normalized attention, a value of γ\gamma, or a measurement. Fig. 5b plots errors against ground truth rather than a fixed-point residual: if one model was iterated longer (Q10), concentration MSE changing about 48x from 10 to 15 iterations (then x0.46 by 20) suggests the iterates are far from converged, which Sec. 4.3's explanation does not account for. How was convergence verified?
  4. How is "theoretically supports infinitely large domains without global resolution degradation" (Sec. 2) consistent with rescaling coordinates to [0,π/2][0, \pi/2] over the domain (App. C.1), normalizing periodic displacement features by the domain length (Eq. 81), and averaging attention over all NN particles, which make the operator depend on domain size and NN?

Addenda

  1. How accurate is the model at the particle counts behind the 38x figure (peak inference memory; Fig. 3b, Sec. 4.2)? Above 2,500 particles the only accuracy evidence is qualitative. Sec. 4.2 omits 3D comparisons for lack of scale-matched baselines, but this does not preclude providing measurements of error with regard to ground truth.
  2. Thm. A.2 covers only the source-free update, and the limiter scales FijF_{ij} but not the source ϕ\phi. What are M−M_-, r−r_- and the elemental-balance residuals in the 3D reactive runs?

Informativeness

Clarifications

  1. The physical rationale for the global flux term, "nonlocal momentum and pressure interactions" (Sec. 4.3), is reasonable for incompressible flow. Yet no ablation removes ConsAttn, so only its nonlocality is measured, against a 34x effect of removing MeshConv (Table 2c). What does attention contribute in total, and is pressure coupling the intended motivation?
  2. The limiter follows Zalesak's construction with global bounds: αi=min⁡{1,(mi−mmin⁡)/(τiOi),(mmax⁡−mi)/(τiIi)}\alpha_i = \min\{1, (m_i - m_{\min})/(\tau_i O_i), (m_{\max} - m_i)/(\tau_i I_i)\} (Sec. 3.2) merges his two nodal factors and limits the whole flux rather than an antidiffusive correction, and where he takes the edge-wise minimum of the endpoint factors, this paper takes their product, which factors and so avoids particle pairs (Eq. 38). What does the product cost in over-limiting, and why is flux-corrected transport (Boris & Book, 1973; Zalesak, 1979) not discussed?
  3. Only the WPW_P-projected latent directions carry physical meaning. Beyond sharing one transport across KK projections (Remark 3.1), what does conserving in latent space buy over predicting skew-symmetric physical fluxes directly?
  4. Sec. 3.1 motivates the implicit form by stability and accuracy at large Δt\Delta t in settings with sources. Was an explicit update unstable or inaccurate at the Δt\Delta t used here?
  5. How would the framework handle open boundaries, where conserved quantities cross ∂Ω\partial\Omega?

Addenda

  1. App. B identifies the attention kernel's first and second moments with an advection velocity and a numerical viscosity. Do the trained kernels match the actual flow this way?
  2. App. B.3 predicts a "deadlock" under symmetric gating, yet in Table 2a, single-sided and symmetric limiting give the same Lconc\mathcal{L}_{\text{conc}} to four digits. Does the deadlock never occur, do the App. B distinctions have no measurable effect, or is the table in error? Where and how often does the limiter activate?
  3. Sec. 3.3 and Remark 3.2 motivate MeshConv by a low/high frequency division of labor with attention. Do the trained modules actually divide the spectrum that way?
  4. In each branch, what do the learned flux kernels look like as functions of distance and direction relative to the flow?
  5. The skew-symmetrized ConsAttn flux matrix has rank at most twice the feature dimension. For which flow features does this become limiting?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.