acceptodds
Under review as a conference paper at ICLR 2027

Mesh-Admissible State Updates in Graph Neural Simulators via Mesh–GENERIC Compatibility

Abstract

Graph neural simulators (GNSs) learn state evolution on discretized physical domains, but trajectory accuracy alone does not certify that a learned update is physically admissible . Structure-preserving methods already incorporate physical and geometric structure into graph - and mesh-based simulators . We ask when the GENERIC reversible–dissipative algebraic contract can hold on a prescribed mesh state space, with its metric, admissible directions, and boundary ports governing the executed dynamics . We call this condition Mesh-GENERIC Compatibility (MGC) and give a constructive sufficient condition that enforces skew adjointness, positive semidefiniteness, and energy–entropy degeneracy for every admissible neural readout, independent of training . MAGNS realizes this construction through neural dynamical assembly, which maps learned constitutive and operator quantities into the MGC-compatible operator family before numerical integration. Controlled interventions test integration, boundary exchange, operator assembly, and Hodge-space admissibility. On long-horizon acoustics, MAGNS matches the strongest structured baseline in median rollout accuracy while reducing mean rollout MSE by 6.3% and attaining a smaller common-chart energy gap. Darcy interventions show that changing the declared Hodge space controls harmonic circulation with fixed neural weights, while errors in the remaining flux components are nearly unchanged.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural simulators, structure-preserving learning, GENERIC / metriplectic dynamics, discrete exterior calculus, Hodge decomposition

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

This paper introduces Mesh–GENERIC Compatibility (MGC), which combines mesh geometry with the algebraic conditions of reversible–dissipative dynamics. The proposed MAGNS framework learns energy, entropy, and operator coefficients and assembles them into compatible operators before numerical integration. The construction guarantees admissible bulk directions, skew-adjointness, positive semidefiniteness, and energy–entropy degeneracy under the stated assumptions. The experiment results demonstrate competitive rollout accuracy and selective control of circulation through changes in the mesh declaration.

Critical strengths
  1. The channel assembly provides a constructive certificate that enforces mesh admissibility and the GENERIC algebraic conditions. The paper also carefully distinguishes these guarantees in continuous time from the properties of numerical integration.
Critical weaknesses
  1. Figure 1 feels somewhat disjointed, making the overall method difficult to follow.

  2. Although Appendix B provides asymptotic cost estimates, the paper does not report runtime or memory comparisons with the baselines. This makes the practical efficiency of MAGNS difficult to assess.

Questions for authors
  1. What practical advantages does the supported-completion construction offer over a GENERIC parameterization built directly within the declared admissible subspace, using the induced Hodge metric?

  2. What are the computational trade-offs of MAGNS relative to existing graph neural simulators, particularly when scaling to larger meshes?

5 comments
AuthorsOpenReview ↗

Response to Reviewer aTNg:Critical Weaknesses W1-W2

Dear Reviewer aTNg

We appreciate your recognition of the constructive compatibility certificate. We have clarified the computation and the role of supported completion, and added runtime and memory measurements.

Critical Weaknesses

W1. The computational flow in Figure 1

Figure 1 feels somewhat disjointed, making the overall method difficult to follow.

We apologize that the original figure left the relationship between its branches unclear, and appreciate your pointing this out. We have redrawn Figure 1 and revised its caption (p. 2) to follow constitutive learning → compatible assembly → time integration, so the reader can trace how the readouts produce a state update.

1. Constitutive learning: what the network predicts. The network reads energy EE, entropy SS and operator coefficients K,RK,R from the shared representation χ\chi. These quantities enter operator assembly; the mesh declaration supplies the metric, admissible directions and completion supports.

2. Compatible assembly: how the readouts determine the dynamics. The entropy gradient determines the supported lift ASA_S for the reversible channel; the energy gradient determines AEA_E for the dissipative channel. We assemble

M=\sum_b A_{E,b}R_b A_{E,b}^{\dagger}.$$ The coefficient blocks $K_a$ are skew-adjoint and $R_b$ are positive semidefinite. The lifts lie in the declared admissible space and enforce $L\nabla S=0$ and $M\nabla E=0$. The assembled $L$ acts on the energy gradient and $M$ on the entropy gradient, giving $L\nabla E+M\nabla S$. The gradients therefore have two roles: they impose orthogonality during lift construction and drive motion when the operators are applied. The revised diagram labels these relationships at the supported lifts, operator factorizations and continuous vector field. **3. Time integration: from vector field to trajectory.** Boundary ports enter the vector field before the integrator advances it to $z_{t+1}$. The next state returns to the input for the following frame, completing the autoregressive loop. Q1 explains how the supported lifts specify motion directions and energy reception. ### **W2. Measured runtime and memory** > Although Appendix B provides asymptotic cost estimates, the paper does not report runtime or memory comparisons with the baselines. This makes the practical efficiency of MAGNS difficult to assess. We appreciate your request for measured costs, and apologize for omitting them from the original manuscript. **For structured dynamical GNS, runtime and memory depend on the constitutive representation, operator execution, integrator and substep count.** Neural functionals require differentiation; substeps, implicit solves and recomputation also affect accuracy, time and memory. Our original Appendix B therefore focused on assembly's asymptotic cost, separating it from these constitutive and numerical choices. We now complement that analysis with end-to-end measurements under specified configurations on a larger CFD mesh. **Native $128\times128$ grid, comprising 16,384 spatial cells.** We compare MAGNS with neural functionals, MAGNS with static learned coefficients, and MeshFT. All use effective batch size eight and 4,500 updates, followed by $H=20$ evaluation on the same validation set. | Method/constitutive form | H20 MSE | Full training time (h) | Peak training memory (GiB) | |---|---:|---:|---:| | MAGNS, neural $E/S$ functionals | 0.23157 | 7.97 | 19.60 | | MAGNS, static learned coefficients | 0.23139 | 3.78 | 2.59 | | MeshFT | 0.24498 | 2.45 | 0.554 | Each entry is a complete run with the same training seed, and all validation windows finish. Runtime covers the full training budget; memory is peak allocated CUDA memory. MAGNS uses Heun and MeshFT uses exponential midpoint splitting, each with one substep per frame. The comparison thus includes each model and its integrator. **Controlled constitutive measurement.** On the same CFD grid, holding training inputs, effective batch size eight and Heun fixed, static coefficients reduce mean time over the first ten $H=1$ updates from 2.797 to 1.662 s, and peak memory from 4.949 to 1.162 GiB. This fixes the inputs and time-stepping method to isolate constitutive cost, while the table captures resource use over the full curriculum. Both MAGNS forms use the same compatible assembly and achieve prediction errors near 0.23 over the complete validation set. For this CFD task, static learned coefficients reduce training cost while retaining compatibility and comparable accuracy. MeshFT provides the structured-baseline comparison on the same task. We will include these configurations, errors and measured costs in the revision. Q2 uses controlled thermoelastic and electromagnetic measurements to explain where the costs arise.
AuthorsOpenReview ↗

Response to Reviewer aTNg:Q1

Q1. Supported completion versus a direct induced-subspace construction

What practical advantages does the supported-completion construction offer over a GENERIC parameterization built directly within the declared admissible subspace, using the induced Hodge metric?

We appreciate your question about the construction's practical role, and apologize for leaving its relation to direct subspace parameterization unclear. Supported completion places thermodynamic compensation on chosen local supports, giving each channel a motion direction and an exchange location.

1. The induced metric and the completion support serve different purposes. For a basis UU of the declared admissible space D\mathcal D, a direct construction can use U⊤GUU^\top G U to define adjoints and orthogonality, then impose energy–entropy degeneracy. This yields a valid compatible parameterization. With full-range lifts and unrestricted coefficient blocks, Proposition 1 characterizes the same pointwise compatible operator class. A finite model also needs local channels and their couplings. The induced metric defines inner products and adjoints; supported completion constructs directions for those channels.

2. Retain the primary motion and specify where its compensation is received. In a GG-orthogonal state split Dx⊕Dy\mathcal D_x\oplus\mathcal D_y, write g=(gx,gy)=∇GFg=(g_x,g_y)=\nabla_G F, where F=SF=S for reversible channels and F=EF=E for dissipative channels. On a receiving support with a nonzero component of gyg_y, choose rar_a such that

P_xC_au \\\\ -r_a\langle g_x,P_xC_au\rangle_G \end{pmatrix},\qquad \langle g_y,r_a\rangle_G=1.$$ The first block retains the selected primary motion; the second cancels its pairing with $g$ on the chosen support, giving $\langle g,A_{F,a}u\rangle_G=0$. Already orthogonal channels need no compensation. For dissipation converted to internal heat, the supports specify which thermal degrees of freedom receive energy, while learned coefficients determine the exchange strength. Channels with different receiving locations can be combined while preserving the algebraic conditions. A direct subspace construction with the same complete local direction family can represent the corresponding operators; supported completion gives a concrete construction of that family. **3. The electromagnetic–thermal experiment tests prediction and local reception.** It combines propagation, temperature-dependent dissipation and local heat reception, without thermal conduction. We report MAGNS, three baselines and three controls. Free op removes direction binding; PostRestrict applies the restriction after full-space assembly; the single-Riesz control replaces local receiving directions with one global thermal direction. All are evaluated on the same validation set at $H=200$. MSE uses common field scales; entries are mean $\pm$ SD. | Method/control | Total MSE | Local thermal-state MSE | |---|---:|---:| | MAGNS (Full) | $(1.57\pm0.16)\times10^{-5}$ | $(1.13\pm0.68)\times10^{-5}$ | | MeshFT | $(3.72\pm0.80)\times10^{-4}$ | $(8.62\pm2.36)\times10^{-4}$ | | Local-TIGNN | $0.01023\pm0.00343$ | $0.02584\pm0.00957$ | | MGN | $0.96311\pm0.04540$ | $0.41296\pm0.14762$ | | free op | $(3.68\pm4.45)\times10^{-4}$ | $(7.86\pm10.96)\times10^{-4}$ | | PostRestrict | $(2.13\pm1.57)\times10^{-5}$ | $(2.43\pm2.34)\times10^{-5}$ | | Subspace, one thermal Riesz direction | $0.35055\pm0.00331$ | $0.95326\pm0.01122$ | The baselines compare prediction performance; the controls change direction binding, assembly order and receiving support. For Full, free op, PostRestrict and the single-Riesz control, the window-mean maximum relative learned-energy drifts are $6.72\times10^{-10}$, $6.71\times10^{-10}$, $1.11\times10^{-8}$ and $1.78\times10^{-3}$, respectively. Every sampled entropy increment is positive in all four groups. The receiving-support difference is clearest in local thermal error. On one training state evaluated with the trained model, the reference frame-averaged thermal rate has relative norm $0.306$ outside the single Riesz direction, whereas the control's corresponding component is only $2.49\times10^{-16}$. This single-Riesz support omits spatial directions present in the reference heating rate, even when algebraic degeneracy holds. Local supported completion retains receiving directions through which the spatial distribution of heat can be learned. The full comparison and mechanism analysis now appear in Section 5.5 and Appendix D.6.
AuthorsOpenReview ↗

Response to Reviewer aTNg:Q2

Q2. Computational trade-offs and larger meshes

What are the computational trade-offs of MAGNS relative to existing graph neural simulators, particularly when scaling to larger meshes?

We appreciate your request to explain computational cost and mesh scaling, and apologize for the incomplete discussion. W2 reports costs on the 16,384-cell CFD grid. The controls below separate the costs of constitutive learning, compatible dynamics execution and time integration.

1. Constitutive flexibility versus differentiation cost. State-dependent neural E/SE/S functionals reduce reliance on a prescribed analytic form for unknown or complex coupled relations. The network can change the functional's shape across states, but training differentiates through its state derivatives. When a chosen functional family is adequate, one can instead learn coefficients from fixed geometry or material conditions and compute state derivatives analytically. The coefficients remain fixed during rollout, while energy, entropy and dynamics vary with the state. The coefficients are learned, L/ML/M may remain state-dependent, and both forms use the same compatible assembly.

A 128-node Timoshenko–Fourier thermoelastic control uses shared constitutive priors, fixed inputs, one-step Heun and batch size one. Replacing neural functionals with static coefficients reduces H=20H=20 forward/backward time from 6.35 to 1.78 s, H=200H=200 rollout from 35.73 to 7.07 s, and training peak memory from 695.14 to 91.87 MiB. As in the CFD example in W2, changing the constitutive form changes differentiation cost while preserving the assembly construction.

2. Per-call cost must be read with accuracy and structural capability. On the 512-cell electromagnetic–thermal task, MAGNS and MeshFT both use RK4 with two substeps per frame, hence eight RHS calls. Same-GPU measurements with matched inputs, float64 and batch size one give H=200H=200 rollout times of 37.47 and 9.27 s, and training peak memory of 1,738 and 115 MiB. Each RHS call includes network computation, functional differentiation and operator application, so equal call counts can still give different runtimes.

The final three-seed results for this task in Q1 give total MSEs of 1.57×10−51.57\times10^{-5} and 3.72×10−43.72\times10^{-4}, and local thermal-state MSEs of 1.13×10−51.13\times10^{-5} and 8.62×10−48.62\times10^{-4}, for MAGNS and MeshFT respectively. The additional computation thus accompanies approximately 95.8% lower total error and 98.7% lower local thermal error, while retaining MAGNS's declared direction and energy–entropy compatibility. MGN's direct increment avoids multistage execution; its prediction results are in Q1. Model choice depends on this cost difference together with the required accuracy and structural properties. MGC fixes the compatibility conditions while leaving the constitutive form, network size and integrator adjustable. The static coefficients in W2 are one way to reduce cost.

3. Changing the integrator for the same model. Holding weights, initial state, mesh and forecast interval fixed, a 512-node MeshFT thermoelastic model gives H=20H=20 MSE 0.08393290.0839329 with either RK4 at eight substeps per frame or adaptive RK45, using 640 and 980 RHS evaluations, respectively. A separate control fixes a MAGNS model and one H=4H=4 window. RK4 with four substeps per frame is unstable; implicit midpoint at the same step size gives MSE 0.0014540.001454 and maximum relative learned-energy drift 5.52×10−125.52\times10^{-12}, using 592 RHS evaluations. The first control reaches similar accuracy with different RHS counts; the second obtains stable integration through additional nonlinear solves. Runtime therefore depends on the method, substep count and cost of each RHS call.

4. Two effects of mesh refinement. More nodes and edges increase message passing, activation storage and channel application. Frequencies resolved by the spatial discretization also affect explicit stability limits and implicit iterations. MAGNS applies factorized operators without forming dense state-space L/ML/M. A rank-rar_a channel on nan_a coordinates costs O(nara+ra2)O(n_ar_a+r_a^2) with a dense coefficient block, plus projection and functional differentiation. Mesh refinement can therefore increase both the cost of one operator application and the work needed for time integration. Generalization across resolutions also requires consistent approximations on different discretizations. Combining compatible dynamics execution with graph operators offers a way to study physical constraints and discretization consistency together.

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

The paper proposes MAGNS, a graph neural simulator that jointly enforces mesh-admissible dynamics and GENERIC compatibility. Channel lifts place reversible and dissipative directions in D∩(∇S)⊥\mathcal{D}\cap(\nabla S)^\perp and D∩(∇E)⊥\mathcal{D}\cap(\nabla E)^\perp, respectively, while operator assembly enforces skew-adjointness, positive semidefiniteness, and energy–entropy degeneracy. The paper establishes continuous-time balance identities and characterizes the full compatible pointwise operator class under full-range lifts and unrestricted coefficient blocks. Experiments on waves, acoustic scattering, Beam3D, and Darcy flow examine prediction accuracy, numerical integration, and fixed-weight control of admissible Hodge sectors.

Critical strengths

The construction addresses a real compatibility issue: restricting an otherwise thermodynamically compatible operator to mesh-admissible directions can destroy its degeneracy conditions. Assembling directly within the intersection resolves this cleanly. The pointwise characterization is particularly appealing: with full-range lifts and unrestricted coefficient blocks, the construction recovers the entire compatible PSD cone and skew-adjoint operator space.

Compatibility is enforced by the architecture independently of trajectory fitting. The fixed-weight Darcy intervention gives this property a concrete use: changing the declared Hodge sectors selectively controls harmonic motion while leaving the remaining components largely unaffected. I find this separation between learned constitutive behavior and prescribed admissibility useful.

Critical weaknesses

The main empirical gap is the absence of a genuinely joint test of the proposed compatibility mechanism. Waves and acoustics use M=0, Beam3D activates both operators but uses the full direction space, and the nontrivial Hodge restriction is tested only in purely dissipative Darcy flow. These experiments validate useful specializations, but do not establish how the construction behaves when geometric restriction and reversible–dissipative coupling are simultaneously consequential.

The predictive benefit of the specific compatibility constraints remains less clear than their algebraic benefit. In the single-seed Beam3D ablation, full assembly and equal-rank free operators have similar mean errors, although full assembly improves the hardest trajectory. The modest acoustic improvement also uses substantially more integration substeps than MeshFT. Stronger controls would help determine when exact compatibility improves simulation accuracy.

Questions for authors
  1. Could the authors add a coupled experiment with PD=I−ΠharP_{\mathcal{D}}=I-\Pi_{\mathrm{har}} and both L∇EL\nabla E and M∇SM\nabla S making non-negligible contributions? The excluded harmonic sector should be active in the unrestricted control, so that the restriction is consequential. Comparing compatible reassembly against imposing the mesh restriction after full-space GENERIC assembly, with shared constitutive readouts and matched integration settings, would directly test the central mechanism. Please report degeneracy residuals, energy–entropy balance errors, and rollout accuracy.

  2. Could the authors repeat the Beam3D full-assembly versus equal-rank free-operator ablation across seeds, and provide an acoustic comparison with matched integration budgets or a convergence study? These controls would clarify how much predictive benefit comes from the compatibility constraints, given the similar mean errors in the current Beam3D ablation and the different acoustic substep counts.

5 comments
AuthorsOpenReview ↗

Response to Reviewer sbHY:Critical Weaknesses W1-W2

Dear Reviewer sbHY

We appreciate your suggestions for testing joint compatibility and controlling numerical resolution. We address the coupled mechanism and the effect of acoustic integration resolution separately.

Critical Weaknesses

W1. A joint test of geometric and thermodynamic compatibility

The main empirical gap is the absence of a genuinely joint test of the proposed compatibility mechanism. Waves and acoustics use M=0, Beam3D activates both operators but uses the full direction space, and the nontrivial Hodge restriction is tested only in purely dissipative Darcy flow. These experiments validate useful specializations, but do not establish how the construction behaves when geometric restriction and reversible–dissipative coupling are simultaneously consequential.

We apologize for the gap in the original experiments and appreciate your call for a joint test. We added a Timoshenko–Fourier thermoelastic coupling experiment to test nontrivial geometry and reversible–dissipative coupling jointly. The system couples five spatial fields to a global thermal reservoir on a periodic mesh.

The declaration PD=I−ΠharP_{\mathcal D}=I-\Pi_{\rm har} excludes the spatially constant mode only in the temperature-rate block; other blocks remain unchanged. In common normalized coordinates, the mean norms of the reversible and dissipative actions L∇EL\nabla E and M∇SM\nabla S are 33.1333.13 and 1.501.50. Both channels are active. The excluded-direction action averages 0.2810.281 in free op and 2.39×10−162.39\times10^{-16} in Full, so the restriction controls a direction that the unrestricted model actually uses.

With matched constitutive parameterization, training and integration conditions, the three-seed H=200H=200 MSE falls from 0.01471±0.001060.01471\pm0.00106 for free op to 0.01102±0.001780.01102\pm0.00178 for Full (25.1%). Temperature-mean drift affects energy allocation through coupling to the global reservoir ee; correspondingly, thermal-reservoir error falls by 51.1%.

Full has maximum degeneracy residuals ∥L∇S∥=6.79×10−17\|L\nabla S\|=6.79\times10^{-17} and ∥M∇E∥=0\|M\nabla E\|=0, with positive sampled internal entropy production. Its common physical-energy error is 2.049%2.049\%, versus 2.096%2.096\% for free op. The direction restriction improves reservoir prediction even though the total-energy errors are similar.

W2. Prediction gains, repeated training and acoustic substeps

The predictive benefit of the specific compatibility constraints remains less clear than their algebraic benefit. In the single-seed Beam3D ablation, full assembly and equal-rank free operators have similar mean errors, although full assembly improves the hardest trajectory. The modest acoustic improvement also uses substantially more integration substeps than MeshFT. Stronger controls would help determine when exact compatibility improves simulation accuracy.

We apologize for not clearly separating the effects of structural constraints and numerical resolution, and appreciate your request for stronger controls. We examine these two effects separately below.

Predictive benefit of compatibility constraints. In Beam3D, PD=IP_{\mathcal D}=I, so removing the geometric restriction leaves the admissible space unchanged. Three-seed MSEs are 0.035786±0.0074150.035786\pm0.007415 for Full and 0.035791±0.0069980.035791\pm0.006998 for free op, with zero geometric leakage for both. The close errors are consistent with this trivial restriction.

This is why the original Darcy ablation tests selective control of harmonic flow through a nontrivial direction restriction, with fixed weights. It activates only dissipation. Timoshenko–Fourier thermoelastic coupling combines nontrivial restriction with both active channels. With matched constitutive parameterization, training and integration, Full lowers error by 25.1% against free op and 34.4% against PostRestrict, which restricts after full-space assembly, while preserving geometry and energy–entropy degeneracy.

Acoustic integration substeps. Holding each model's weights fixed, we evaluated both methods on the same set of validation windows using KDK42 at H=80H=80. MAGNS and MeshFT obtain MSEs of 0.0157210.015721 and 0.0215030.021503, respectively, a 26.9% reduction for MAGNS.

With MeshFT weights and evaluation windows fixed, refinement from 8 to 21 and 42 substeps changes MSE by only 0.0215% overall. Its eight-substep result is already close to the finer-resolution results, and refinement does not remove the MAGNS advantage.

The updated multi-seed results also retain this advantage: MAGNS reduces the aggregate validation and test MSEs by 28.0% and 14.0%, respectively, relative to MeshFT.

AuthorsOpenReview ↗

Response to Reviewer sbHY: Q1

Q1. Active harmonic exclusion and the order of compatible assembly

Could the authors add a coupled experiment with PD=I−ΠharP_{\mathcal D}=I-\Pi_{\rm har} and both L∇EL\nabla E and M∇SM\nabla S making non-negligible contributions? The excluded harmonic sector should be active in the unrestricted control, so that the restriction is consequential. Comparing compatible reassembly against imposing the mesh restriction after full-space GENERIC assembly, with shared constitutive readouts and matched integration settings, would directly test the central mechanism. Please report degeneracy residuals, energy–entropy balance errors, and rollout accuracy.

We apologize for the incomplete control comparisons and appreciate this concrete proposal. For Timoshenko–Fourier thermoelastic coupling, we compare MAGNS, MeshFT, Local-TIGNN and MGN, together with free op and PostRestrict.

1. Nontrivial restriction and active channels. Five spatial fields couple to a global thermal reservoir ee on a periodic mesh. PD=I−ΠharP_{\mathcal D}=I-\Pi_{\rm har} excludes the constant mode only from the temperature-rate block; other blocks remain unchanged. Full has mean action norms 33.1333.13 and 1.501.50 for L∇EL\nabla E and M∇SM\nabla S. In free op, excluded-direction action and temperature harmonic MSE are 0.2810.281 and 0.004720.00472; Full reduces them to 2.39×10−162.39\times10^{-16} and 4.06×10−324.06\times10^{-32}. Both channels are active, and the restriction removes a direction used by the unrestricted model.

2. Fixed readouts isolate assembly order. We hold the state and constitutive readouts fixed and change only the order of completion and restriction. Compatible reassembly gives maximum ∥L∇S∥=1.01×10−16\|L\nabla S\|=1.01\times10^{-16} and ∥M∇E∥=4.33×10−15\|M\nabla E\|=4.33\times10^{-15}; restriction after full-space assembly gives 0.001980.00198 and 0.385480.38548. The rollout results and physical diagnostics below use three independently trained variants.

3. Direction and energy–entropy behavior. For independently trained controls, direction-action norms are sample means and degeneracy residuals sample maxima, in common normalized coordinates.

VariantMean excluded-direction actionMaximum ∥L∇S∥\lVert L\nabla S\rVertMaximum ∥M∇E∥\lVert M\nabla E\rVert
Full2.39×10−162.39\times10^{-16}6.79×10−176.79\times10^{-17}00
free op0.2810.2817.31×10−177.31\times10^{-17}00
PostRestrict3.39×10−163.39\times10^{-16}6.49×10−176.49\times10^{-17}0.428580.42858

The maximum internal learned-energy power ∣E˙θ∣|\dot E_\theta| is 1.18×10−161.18\times10^{-16} for Full and 1.39×10−161.39\times10^{-16} for free op. Although PostRestrict removes geometric leakage, its maximum internal energy power reaches 0.012430.01243. Sampled internal entropy production is positive for all three variants, with a minimum of 2.12×10−32.12\times10^{-3}.

We separate thermodynamic violations from integration error by comparing actual energy and entropy increments with accumulated stage power at each substep. Across the three variants, maximum energy-balance residuals are at most 1.59×10−151.59\times10^{-15} and entropy-balance residuals at most 5.70×10−175.70\times10^{-17}. PostRestrict's nonzero internal energy change therefore comes from violated degeneracy after geometric restriction, not from the integration defect.

4. Long-horizon prediction and three baselines. All methods use common splits, normalized coordinates and metrics. The three structural controls share parameterization, training and integration; only restriction or assembly order changes. Results are mean ±\pm sample SD over three seeds at H=200H=200. Model adaptations and physical diagnostics are given in Appendix D.6 of the revision.

Method or controlTotal MSEThermal-reservoir MSECommon physical-energy error (%)
MAGNS (Full)0.01102±0.001780.01102\pm0.001780.02113±0.005000.02113\pm0.005002.049±0.2842.049\pm0.284
MeshFT0.19508±0.007030.19508\pm0.007030.09056±0.001560.09056\pm0.0015617.832±0.17917.832\pm0.179
Local-TIGNN1.13877±0.128011.13877\pm0.128010.31781±0.111550.31781\pm0.1115550.605±22.93550.605\pm22.935
MGN2.57147±0.237392.57147\pm0.2373911.87184±1.3372411.87184\pm1.337241715.380±838.7691715.380\pm838.769
free op0.01471±0.001060.01471\pm0.001060.04319±0.003410.04319\pm0.003412.096±0.4372.096\pm0.437
PostRestrict0.01681±0.002060.01681\pm0.002060.06140±0.005030.06140\pm0.005033.137±0.7083.137\pm0.708

Physical-energy error uses one reference functional for all methods: framewise relative errors against the true trajectory are averaged over time and trajectories. MAGNS has the lowest means in all three columns. The five-field MSEs for Full and free op are 0.009000.00900/0.009010.00901, but Full lowers reservoir error by 51.1%. Restricting the constant temperature-rate mode controls temperature-mean drift and its coupling to the global reservoir ee, so the gain is concentrated in coupled energy allocation.

AuthorsOpenReview ↗

Response to Reviewer sbHY: Q2

Q2. Repeated Beam3D controls and acoustic numerical refinement

Could the authors repeat the Beam3D full-assembly versus equal-rank free-operator ablation across seeds, and provide an acoustic comparison with matched integration budgets or a convergence study? These controls would clarify how much predictive benefit comes from the compatibility constraints, given the similar mean errors in the current Beam3D ablation and the different acoustic substep counts.

We apologize for the limited original controls and appreciate your request for repeated runs and matched numerical comparisons. We repeated the Beam3D comparison, evaluated the acoustic models at matched KDK resolution, and examined refinement with a fixed generator.

1. Beam3D: repeated prediction results and direction leakage. The updated three-seed H=20H=20 MSEs are 0.035786±0.0074150.035786\pm0.007415 for Full and 0.035791±0.0069980.035791\pm0.006998 for free op; geometric direction leakage is zero for both. In a same-readout assembly measurement, both variants have reversible entropy-direction leakage of 7.96×10−177.96\times10^{-17} and dissipative energy-direction leakage of 5.03×10−175.03\times10^{-17}. Beam3D uses the complete nodal direction space, PD=IP_{\mathcal D}=I, so removing the geometric restriction leaves the admissible space unchanged. The Timoshenko–Fourier experiment in Q1 tests a consequential restriction: Full reduces error by 25.1% relative to free op while suppressing the active excluded harmonic component.

2. Acoustics: matched temporal resolution preserves the accuracy advantage. Holding each model's weights and the validation windows fixed, MAGNS with KDK21 and Heun21 completes 25% and 45% of the H=80H=80 rollouts, respectively; KDK42 completes all rollouts with MSE 0.0157210.015721. MeshFT completes all rollouts at KDK8, KDK21 and KDK42, with MSEs 0.0215080.021508, 0.0215040.021504 and 0.0215030.021503. Thus, at matched KDK42 resolution MAGNS lowers MSE by 26.9%, while refining MeshFT from eight to 42 substeps changes its MSE by only 0.0215%. The acoustic advantage is therefore not explained solely by the extra integration substeps.

Here 21 and 42 denote substeps per data interval; the prediction horizon remains H=80H=80. Matching KDK42 controls temporal resolution across methods, while refining MeshFT tests whether its eight-substep setting explains the accuracy gap.

3. The CFL baseline and refinement of the learned dynamics. Both protocols use CFL-based substep selection. MeshFT's acoustic protocol specifies a target Courant number of 0.5 and approximately eight substeps per data interval (Noguchi & Kawahara, 2026, Appendix D.5); we retain that setting. The data-grid CFL baseline for MAGNS is

\qquad c_{\max}=2,\quad C_{\rm target}=0.5.$$ The CFL rule sets the initial temporal resolution. We then measure integration error for the learned dynamics. At one validation state on the original grid, we freeze the MAGNS generator and compare propagation with an unsplit matrix-exponential reference. The one-frame relative discrepancies are $227.913$ for KDK21 and $0.05971$ for KDK42. With the generator fixed, the lower discrepancy at this state measures the effect of numerical refinement. The matched-resolution rollout comparison above tests completion and prediction accuracy. These tests answer different questions: matched-resolution rollouts test whether more substeps explain the advantage; the frozen-generator comparison tests whether refinement reduces MAGNS integration error. The repeated-training results agree with these controls: MAGNS reduces the aggregate MSE relative to MeshFT by **28.0% on VALID** ($0.01670\pm0.00356$ vs. $0.02319\pm0.00072$) and **14.0% on TEST** ($0.02002\pm0.00271$ vs. $0.02328\pm0.00017$).
Reviewer_yj1aOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper extends graph neural networks for mesh-based simulation to satisfy two types of structure at the same time:

  • Thermodynamic structure (GENERIC), which writes the dynamics as a reversible part driven by the energy gradient plus a dissipative part driven by the entropy gradient. Two degeneracy conditions are added to guarantee that a closed system conserves energy and its entropy never decreases.

  • Mesh structure: here, the flow is restricted to a task-declared subset of the following directions: gradient-like (exact), curl-like (coexact), and global circulation around holes in the domain (harmonic). The declaration also fixes the mesh metric and the boundary ports through which energy can enter or leave.

Satisfying both constraints on the same learned dynamics is called Mesh–GENERIC Compatibility (MGC). The presented approach, MAGNS, uses a graph encoder–processor similar to MGN, but instead of decoding the next state it predicts energy and entropy functionals plus operator coefficients. From these it assembles operators that satisfy MGC by construction, for any network output, and passes the resulting vector field to a numerical integrator.

The approach is competitive with or better than structured baselines (MeshFT, IGNS, Local-TIGNN, MGN) on wave propagation, acoustic scattering and a viscoelastic beam.

Critical strengths
  • new formulation that combines two phsyics informed constraints to the prediction of the mesh dynamics
  • the approach outperforms SOTA in most cases
  • experiments are exhaustive and cover a diverse set of dynamics
Critical weaknesses
  • The ablation "free operator" is less clean than claimed with very small differences in the results. So most of the gain comes from “learn a vector field via E, S, L, M rather than an increment”; the strict compatibility constraints add a quite small, case-dependent benefit.
  • Many experiments use single seeds results, statistical signifcance is therefore questionable
Questions for authors
  • can you run the single seed experiments at least for 3 (better 5) seeds?
  • The paper’s guarantee holds for the model’s idealized continuous-time dynamics and for its own learned notion of energy. In practice, though, simulations run in discrete time steps and are judged against the real physical energy. In the Darcy experiment, for example, the learned energy is conserved almost perfectly, while the physical energy drifts by about 1.7%. Could the authors explain how much of the guarantee survives in practice, and show whether the version without the guarantee (“free operators”) behaves noticeably worse on energy, not just on prediction error?
3 comments
AuthorsOpenReview ↗

Response to Reviewer yj1a:Critical Weaknesses W1-W2

Dear Reviewer yj1a

We appreciate your recognition of the joint formulation and the range of dynamics studied. We have added repeated runs and coupled-system experiments to assess prediction accuracy, geometric constraints and thermodynamic compatibility.

Critical Weaknesses

W1. Predictive benefit of the compatibility constraints

The ablation "free operator" is less clean than claimed with very small differences in the results. So most of the gain comes from “learn a vector field via E, S, L, M rather than an increment”; the strict compatibility constraints add a quite small, case-dependent benefit.

We apologize for not explaining the scope of this ablation clearly and appreciate your request to isolate the benefit of compatibility.

The small Beam3D difference is consistent with its trivial Hodge-direction restriction. In this setting, the declaration uses the full nodal direction space, PD=IP_{\mathcal D}=I, so removing the direction restriction leaves the admissible space unchanged. This motivated the Darcy experiment in the submitted paper, which tests a nontrivial Hodge-sector declaration.

Across three Beam3D training runs, Full and free op obtain 0.035786±0.0074150.035786\pm0.007415 and 0.035791±0.0069980.035791\pm0.006998, respectively. Both have zero geometric direction leakage under this declaration. With the same constitutive readouts, both assemblies give maximum normalized thermodynamic direction leakage of 7.96×10−177.96\times10^{-17} for the reversible entropy pairing and 5.03×10−175.03\times10^{-17} for the dissipative energy pairing. Their similar prediction errors come with similar structural behavior.

Darcy tests whether the declaration selectively controls a nontrivial direction. On the periodic domain, case A contains physical harmonic circulation, whereas case B has none. Holding the trained weights fixed, we switch the execution declaration from II to I−ΠharI-\Pi_{\rm har} and reassemble the compatible mobility. In case B, the harmonic-flux error falls from 3.33×10−53.33\times10^{-5} to 4.03×10−184.03\times10^{-18}. The same switch also removes the physical harmonic component in case A, while errors in the remaining flux directions change by less than 0.5%0.5\% in both cases. The switch controls the selected global mode while leaving the other flow components almost unchanged.

The new Timoshenko–Fourier thermoelastic coupling experiment extends this test to coupled reversible and dissipative dynamics. Both channels are active, with average action norms of 33.1333.13 and 1.501.50 in common normalized state coordinates. Full and free op use the same constitutive parameterization, training and integration conditions; the controlled difference is the Hodge-direction restriction. Here I−ΠharI-\Pi_{\rm har} acts only on the temperature-rate block, excluding its spatially constant mode while retaining the other state blocks. This directly constrains temperature-mean drift and affects energy allocation through the coupling to the global thermal reservoir ee.

Across three independently trained seeds, their H=200H=200 MSEs are 0.01102±0.001780.01102\pm0.00178 and 0.01471±0.001060.01471\pm0.00106, a 25.1% reduction for Full. The excluded-direction action averages 0.2810.281 for free op and 2.39×10−162.39\times10^{-16} for Full. Correspondingly, reservoir ee error decreases from 0.043190.04319 to 0.021130.02113, while the five-field errors are 0.009010.00901 and 0.009000.00900. With the E,S,L,ME,S,L,M learning formulation held fixed, this comparison demonstrates an additional benefit from the direction constraint when the excluded sector is active.

W2. Coverage across independent training seeds

Many experiments use single seeds results, statistical signifcance is therefore questionable

We apologize for the limited seed coverage in the original submission and appreciate your request for repeated runs. We have added the comparisons below to assess stability.

The Beam3D structural ablations now report three training runs per variant, covering Full, free op, No assembly and no completion.

The new Timoshenko–Fourier thermoelastic coupling and electromagnetic–thermal coupling experiments each compare four methods: MAGNS, MeshFT, Local-TIGNN and MGN. Adding free op and PostRestrict gives six configurations for Timoshenko–Fourier. Electromagnetic–thermal coupling also includes a single-Riesz receiving-support control, giving seven configurations. We report means and sample standard deviations across training seeds.

The existing wave, Beam3D main comparison and Darcy experiments retain their five-, three- and three-seed coverage, respectively.

AuthorsOpenReview ↗

Response to Reviewer yj1a: Q1

Q1. Repeating the single-seed experiments

can you run the single seed experiments at least for 3 (better 5) seeds?

We apologize for the limited repetitions in the original submission and appreciate your suggestion to add seeds. The acoustic and Beam3D results below report means and sample standard deviations.

Acoustic prediction. The table summarizes the original and additional results, with all methods evaluated on the same test samples.

MethodVALID rollout MSE, H=80H=80TEST rollout MSE, H=80H=80
MAGNS0.01670±0.003560.01670\pm0.003560.02002±0.002710.02002\pm0.00271
MeshFT0.02319±0.000720.02319\pm0.000720.02328±0.000170.02328\pm0.00017
IGNS0.04874±0.002840.04874\pm0.002840.05011±0.002570.05011\pm0.00257
MGN0.11537±0.002390.11537\pm0.002390.11281±0.000050.11281\pm0.00005

MAGNS reduces mean rollout MSE relative to MeshFT by 28.0% on VALID and 14.0% on TEST.

Beam3D control design and prediction. The controls distinguish the direction restriction, the structured dynamics assembly, and the completion that enforces thermodynamic degeneracy. All prediction results use the same validation trajectories and H=20H=20.

VariantControlled changeRollout MSE ±\pm SD
FullComplete compatible assembly0.035786±0.0074150.035786\pm0.007415
free opRemove the Hodge-direction restriction; retain skew-symmetry, positive semidefiniteness and degeneracy completion0.035791±0.0069980.035791\pm0.006998
No assemblyReplace the functional/operator assembly with a capacity-matched GNN predicting the native state increment0.06861±0.003110.06861\pm0.00311
no completionRetain the functional/operator representation and skew-symmetric/positive-semidefinite construction; remove the completion enforcing L dS=0L\,\mathrm{d}S=0 and M dE=0M\,\mathrm{d}E=00.03028±0.003010.03028\pm0.00301

We have clarified the No-assembly definition and corrected its original table entry. Beam3D uses the full nodal direction space, so after removing structured assembly, direct-prediction models with or without the identity output projection represent the same control. In paired measurements, their predictions and parameter gradients differ by at most 1.90×10−171.90\times10^{-17} and 2.17×10−192.17\times10^{-19}, respectively. We therefore report this control as No assembly. The updated 0.06861±0.003110.06861\pm0.00311 replaces the original 0.17950.1795 entry and its associated 4.53-fold attribution. We apologize for the inconsistent reporting of this control.

Direction leakage. For fixed inputs and constitutive readouts, we compare the three assembly rules using the normalized cross-pairings

ℓrev=∣⟨dS,LdE⟩∣∥dS∥∗∥LdE∥.\ell_{\mathrm{rev}}=\frac{\lvert\langle\mathrm{d}S,L\mathrm{d}E\rangle\rvert}{\lVert\mathrm{d}S\rVert_{\ast}\lVert L\mathrm{d}E\rVert}.

ℓdiss=∣⟨dE,MdS⟩∣∥dE∥∗∥MdS∥.\ell_{\mathrm{diss}}=\frac{\lvert\langle\mathrm{d}E,M\mathrm{d}S\rangle\rvert}{\lVert\mathrm{d}E\rVert_{\ast}\lVert M\mathrm{d}S\rVert}.

The norms use common normalized state coordinates. The table reports the maximum cross-pairing over the evaluated states. Geometric leakage is zero under the full-C0C^0 declaration, PD=IP_{\mathcal D}=I.

Assembly ruleGeometric direction leakageReversible entropy-direction leakage ℓrev\ell_{\mathrm{rev}}Dissipative energy-direction leakage ℓdiss\ell_{\mathrm{diss}}
Full007.96×10−177.96\times10^{-17}5.03×10−175.03\times10^{-17}
free op007.96×10−177.96\times10^{-17}5.03×10−175.03\times10^{-17}
no completion000.500130.500130.051620.05162

Trajectories from the final checkpoints show the same distinction. For Full, the trajectory-maximum norms of L dSL\,\mathrm{d}S and M dEM\,\mathrm{d}E, averaged across runs and trajectories, are 1.34×10−121.34\times10^{-12} and 1.32×10−131.32\times10^{-13}. Without completion, they rise to 4.76×1034.76\times10^3 and 1.17×1031.17\times10^3 in the same state coordinates. The average relative learned-energy imbalance after accounting for external-port work increases from 4.62% to 77.85%. Full's imbalance is accounted for by the discrete-integration defect. For no completion, the defect is only 1.09%; the additional internal energy exchange accounts for most of the imbalance.

The identity direction declaration explains the close Full/free-op predictions and zero geometric leakage discussed in W1. The no-completion control shows that lower prediction error does not imply correct thermodynamic behavior.

AuthorsOpenReview ↗

Response to Reviewer yj1a: Q2

Q2. Discrete-time balance and physical energy fidelity

summary: How do continuous-time guarantees for learned energy carry over to discrete simulation and physical energy? Do free operators perform worse on energy as well as prediction?

We apologize for the unclear distinction and appreciate your request to examine the practical guarantees. MAGNS enforces generator compatibility; finite-step balance depends on the integrator, and physical-energy fidelity on the functional. Free op removes the Hodge-direction restriction but retains thermodynamic completion.

1. Guarantees of the continuous generator. MAGNS satisfies L dSθ=0L\,\mathrm dS_\theta=0, M dEθ=0M\,\mathrm dE_\theta=0, skew-adjointness, positive semidefiniteness and the declared direction restriction at every evaluation, including intermediate integration stages. These identities give E˙θ=0\dot E_\theta=0 and S˙θ≥0\dot S_\theta\geq0 for a closed system; open-system energy change is balanced by boundary-port work.

2. Preservation at finite step size. Runge–Kutta combinations preserve a fixed linear space of bulk directions. Nonlinear energy and entropy balances also depend on the integrator: compatible two-point discrete-gradient assembly satisfies the discrete balances for an exactly solved step; an approximate solve introduces its residual. Heun and ordinary Runge–Kutta stepping leave numerical balance defects.

The same-weight wave experiment illustrates this: replacing Heun with discrete-gradient stepping reduces physical-energy drift from 4.121×10−44.121\times10^{-4} to 2.105×10−52.105\times10^{-5} (94.9%), while trajectory MSE remains of order 10−510^{-5}.

3. Relation to physical energy. Darcy prescribes an independent reference Ephys=12∑ih0,ipi2+∑iVieiE_{\rm phys}=\tfrac12\sum_i h_{0,i}p_i^2+\sum_i V_i e_i. For a closed Full model with a fixed learned functional, define δE=Ephys−Eθ\delta E=E_{\rm phys}-E_\theta. Then

ΔEphys=Rnum+ΔδE,Rnum=ΔEθ.\Delta E_{\rm phys}=R^{\rm num}+\Delta\delta E,\qquad R^{\rm num}=\Delta E_\theta.

The first term is the numerical balance defect; the second is the change in the functional discrepancy along the trajectory. Darcy's maximum relative drifts are 2.03×10−72.03\times10^{-7} for learned energy and 1.67%1.67\% for physical energy. Compatibility prevents spurious internal learned-energy production; physical fidelity also depends on constitutive accuracy and numerical integration. Specifying physical energy directly gives δE=0\delta E=0 while learned energies require independent physical validation.

Full/free-op prediction and energy comparison. Values below are mean ±\pm sample SD for prediction and trajectory energy; H=20H=20 for Beam3D and H=200H=200 for the coupled systems.

SettingMetricFullfree op
Beam3DMSE0.035786±0.0074150.035786\pm0.0074150.035791±0.0069980.035791\pm0.006998
Beam3DNormalized dissipative energy-direction leakage5.03×10−175.03\times10^{-17}5.03×10−175.03\times10^{-17}
Timoshenko–FourierMSE0.01102±0.001780.01102\pm0.001780.01471±0.001060.01471\pm0.00106
Timoshenko–FourierRelative physical-energy error (%)2.049±0.2842.049\pm0.2842.096±0.4372.096\pm0.437
Electromagnetic–thermalMSE(1.568±0.163)×10−5(1.568\pm0.163)\times10^{-5}(3.677±4.445)×10−4(3.677\pm4.445)\times10^{-4}
Electromagnetic–thermalRelative learned-energy drift(6.72±0.09)×10−10(6.72\pm0.09)\times10^{-10}(6.71±0.13)×10−10(6.71\pm0.13)\times10^{-10}

Beam3D. The direction declaration admits the full C0C^0 space (PD=IP_{\mathcal D}=I), so free op leaves the admissible space unchanged. Both have zero geometric leakage and roundoff-level thermodynamic leakage in the same-readout measurement, consistent with their similar prediction errors.

Timoshenko–Fourier thermoelastic coupling. Here I−ΠharI-\Pi_{\rm har} acts only on the temperature-rate block, excluding its spatially constant mode. It targets temperature-mean drift and the coupled thermal reservoir ee. Full lowers MSE by 25.1% and excluded-direction action from 0.2810.281 to 2.39×10−162.39\times10^{-16}. Reservoir MSE falls from 0.043190.04319 to 0.021130.02113 (51.1%), while five-field MSE stays near 0.00900.0090. Similar total-energy errors thus accompany different accuracy in coupled energy allocation.

Electromagnetic–thermal coupling. Both retain very small learned-energy drift (mean window maximum), yet Full lowers MSE by 95.7% and local thermal-energy field MSE from 7.86×10−47.86\times10^{-4} to 1.13×10−51.13\times10^{-5}. Its minimum measured framewise entropy increment is positive, 1.18×10−41.18\times10^{-4}.

The revised Section 5.5 and Appendix D.6 report both coupled systems with MeshFT, Local-TIGNN and MGN.

Removing thermodynamic completion itself. The Beam3D no-completion model has lower MSE (0.03028±0.003010.03028\pm0.00301), but final-checkpoint energy imbalance after external-port work rises from 4.62% to 77.85%. Its integration defect after accounting for internal power is only 1.09%, so the additional internal energy exchange explains most of the imbalance. Lower prediction error does not imply correct thermodynamic behavior.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised manuscript: summary of changes

We appreciate the reviewers’ constructive feedback. We have uploaded a revised manuscript incorporating the additional experiments and analyses discussed in our responses. The main changes are:

  • Presentation and results: Updated Figure 1 and the acoustic and Beam3D ablation results, including the associated physical-behavior analyses.
  • Coupled-system experiments: Incorporated the Timoshenko–Fourier and electromagnetic–thermal comparisons and controls in Section 5.5 and Appendix D.6.
  • Discussion and exposition: Expanded Section 6 to distinguish constitutive learning, structured dynamics execution, and time integration, and to discuss their implications for computational cost. We also revised the abstract and main text for clarity and coherence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.