acceptodds
Under review as a conference paper at ICLR 2027

A-MFO-Net: Boundary-Conditioned GNN–FNO–GNN Operators for Long-Horizon Room-Scale 3D Thermal-Fluid Simulation

Abstract

Long-horizon three-dimensional thermal-fluid simulation on unstructured meshes requires both global information propagation and geometry-sensitive local recon- struction. We introduce A-MFO-Net, a boundary-conditioned GNN–FNO–GNN operator whose graph stages weight edge-conditioned messages using source and destination states and directed edge geometry. On a held-out 22,000-cell extended- room data-center trajectory (90% ACU cooling, 100% server workload), we predict five state channels over 400 steps and 20 seconds. Training is predominantly on 200-step room-scale trajectories, with one shorter extended-room and one longer room-scale case, so the test probes their joint transfer rather than isolated geometry or horizon extrapolation. All seven models use 200 training windows per epoch and seeds 42, 43, and 44. At the final step, A-MFO-Net’s relative-L2 errors (mean ± sample standard deviation) are (6.97 ± 0.18)% for velocity, (31.56 ± 4.75)% for pressure, and (5.05 ± 0.15)% for temperature: 68.6% and 22.8% below the strongest external velocity and pressure comparators, respectively. A-MFO-Net has the lowest mean velocity and pressure errors across all seven models, while GAOT has the lowest mean temperature error. FNO has (39.24 ± 0.99)% velocity error and highly variable pressure error, supporting the implemented graph-containing pipeline but not isolating attention or neighbor aggregation. A-MFO-Net completes the 400-step evaluation in 64.0 ± 1.3 seconds.

Primary area:
generative models
Keywords:
Neural operators, graph neural networks, Fourier neural operators, transient thermal-fluid simulation, data center cooling, mesh-based simulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZVy3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes A-MFO-Net, a graph–Fourier–graph surrogate for autoregressive prediction of velocity, pressure, and temperature on unstructured 33D data-center meshes. It combines local graph processing with global Fourier mixing and weights graph messages using endpoint states and directed geometry. The main contribution is this architectural integration for room-scale thermal-fluid forecasting, with recorded non-interior states restored at each rollout step.

Validation uses 1414 training trajectories and compares 66 alternatives across 33 training seeds on one held-out, 22,00022{,}000-cell trajectory spanning 400400 steps. A-MFO-Net achieves final-step relative L2L_2 errors of 6.976.97 % for velocity, 31.5631.56 % for pressure, and $5.05$ % for temperature, obtaining the lowest mean velocity and pressure errors while GAOT performs best on temperature. Attention-disabled and graph-free controls are included, although they change additional architectural factors and therefore do not isolate individual components’ benefits.

Critical strengths
  1. Substantial velocity-accuracy improvement. Across 33 training seeds, A-MFO-Net achieves a final-step velocity relative L2L_2 error of (6.97±0.18)(6.97 \pm 0.18) %, compared with (22.20±1.17)(22.20 \pm 1.17) % for the strongest external comparator (GAOT), an obvious reduction.

  2. A practically meaningful demonstration of long-rollout, multi-field forecasting. The model jointly predicts velocity, pressure, and temperature over 400400 autoregressive steps on a complete 22,00022{,}000-cell, irregular 33D room mesh under boundary conditioning. This evaluation addresses spatially detailed thermal-fluid forecasting around equipment, giving the work practical significance.

Critical weaknesses
  1. The contribution beyond existing graph–Fourier architectures is insufficiently established. The local–global construction substantially overlaps with GINO, while geometry-aware graph attention also has substantial prior art. The proposed integration could still contribute useful knowledge, but its ablations change message normalization or additional architectural factors, preventing clear attribution of the gains to the proposed components. Consequently, the paper does not convincingly establish a distinct methodological advance or a transferable design insight, which limits its significance for the ML community.

  2. The empirical evidence is too narrow to establish a reliable advantage. All headline results come from one held-out trajectory; the 33 training seeds quantify optimization variability rather than robustness across physical cases. Training already includes the extended-room geometry and a longer source trajectory, so the evaluation demonstrates one particular combination of operating condition, geometry, and duration. Although the authors acknowledge this scope, the lack of independent test conditions leaves the substantial reported improvements insufficiently supported as a broader empirical contribution.

Questions for authors
  1. Scientific contribution. Given the architectural overlap with GINO and related geometry-aware operators, what is the principal new methodological or empirical insight established by this work? Please identify the specific evidence that supports this insight, beyond demonstrating that the integrated architecture performs well on the reported case.

  2. Forecasting task. Which of the restored non-interior variables are independently prescribed inputs, and which are quantities obtained from the CFD solution? Are the reported errors evaluated before or after restoration? These distinctions are necessary to understand what the model actually predicts and whether its conditioning information would be available in the intended application.

  3. Robustness across cases. Are results available on additional held-out operating conditions, or can the advantage be checked through a limited evaluation within the same application? Please distinguish variation across physical cases from variation across training seeds, since this could materially change the assessment of the empirical contribution.

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

The paper proposes A-MFO-Net, a hybrid GNN–FNO–GNN architecture for long-horizon 3D thermal-fluid simulation. Specifically, a graph neural network encodes mesh-cell features, which are then projected onto a regular latent grid for global information processing using an FNO. Another graph neural network subsequently decodes the updated features back to the original mesh. The method also incorporates geometry-aware attention and boundary conditioning. Experiments demonstrate that the proposed method outperforms most baselines on the evaluated test trajectory, achieving the lowest velocity and pressure prediction errors.

Critical strengths
  • The paper provides comparisons with multiple representative baselines, including FNO, MeshGraphNet, GAOT, Transolver++, and FIGConvUNet, using a consistent evaluation protocol.

  • The proposed method achieves superior performance over the baselines in velocity and pressure prediction, with competitive temperature prediction accuracy.

Critical weaknesses
  1. Limited methodological novelty. The proposed GNN–FNO–GNN architecture is highly similar to GINO [1], which also employs graph operators to map features from irregular meshes to a regular latent grid, applies an FNO for global information processing, and maps the resulting features back to the original discretization through another graph operator. Both methods also incorporate signed-distance information to represent geometry. Although the authors acknowledge GINO in the related work, the paper does not sufficiently distinguish the proposed architecture from this closely related method. The authors should clarify the specific advantages of their edge-conditioned attention and boundary-conditioning mechanisms over existing approaches. A quantitative comparison with GINO would also be important to establish the contribution of the proposed method.

  2. Unclear coordinate normalization across different geometries. Section 3.1 introduces normalized coordinates qi\mathbf{q}_i for each mesh cell, but the normalization procedure is not explicitly described. In particular, are the coordinates normalized independently for each domain using its own spatial bounds, or are consistent normalization parameters shared across domains of different sizes? These choices may affect how the model represents geometric scale and transfers information across different room configurations. The authors should clarify the normalization procedure and how the spatial coordinates are used in the mesh-to-grid mapping.

  3. Limited evaluation of generalization and robustness. The evaluation is conducted on only one held-out trajectory, which is insufficient to comprehensively assess the model's robustness and generalization ability. Although experiments are repeated with three random seeds, these only characterize training variability on the same test case. Furthermore, the training dataset already includes an extended-room trajectory and a 400-step room-scale trajectory. Therefore, the held-out test combines previously observed geometry types and temporal horizons under a different operating condition, rather than independently demonstrating generalization to unseen geometries or longer horizons. More diverse test cases and controlled generalization experiments are necessary to establish the broader applicability of the proposed method.

  4. Insufficiently controlled and incomplete ablation studies. Although the main paper reports the attention-disabled MFO-Net baseline in Table 1, the implementation of this ablation is not sufficiently specified. In particular, when attention is disabled, are the coefficients $\alpha_{ij} $ in Equation 2 set to 1 or to 1/∣N(i)∣1/|\mathcal{N}(i)|? These choices are not equivalent because Equation 2 already applies an additional degree-mean normalization, resulting in different scales for the aggregated messages. Consequently, the observed improvement cannot be attributed solely to the learned attention mechanism. Moreover, the current experiments do not systematically investigate several potentially important architectural choices, such as the resolution of the regular latent grid, the number of nearest neighbors, and the respective contributions of the graph encoder and decoder. More carefully controlled and comprehensive ablation studies would strengthen the methodological claims.

  5. Potential limitations in the long-horizon evaluation protocol. As stated in Section 3.3 (lines 193–195), during autoregressive rollout, predictions outside the interior-flow mask are replaced with recorded ground-truth values before the next prediction step. While this is reasonable for prescribed boundary conditions, it means the model continuously receives reference boundary states rather than independently predicting the complete system state. The authors should clarify whether these restored values correspond exclusively to boundary conditions assumed to be known during inference. Furthermore, Equation 3 defines the relative-L2 error over all NN mesh cells. It is unclear whether the evaluation includes cells whose predicted states have been replaced by ground-truth values. If so, these cells could artificially reduce the reported relative error. The authors should clarify the evaluation protocol and provide errors computed exclusively on predicted interior-flow cells.

  6. Lack of evaluation of physical consistency. The experimental evaluation mainly focuses on relative-L2 errors of the predicted velocity, pressure, and temperature fields, without explicitly assessing whether the predicted dynamics satisfy the underlying physical constraints. For example, if the simulated airflow is treated as incompressible, the predicted velocity field should approximately satisfy ∇⋅u=0\nabla\cdot\mathbf{u}=0. Additional evaluations of mass conservation, continuity residuals, and thermal transport consistency would provide stronger evidence that the proposed model produces physically plausible long-horizon simulations, rather than merely achieving low global prediction errors.

References

[1] Li, Z., et al. Geometry-Informed Neural Operator for Large-Scale 3D PDEs. Advances in Neural Information Processing Systems (NeurIPS), 2023.

Questions for authors

Please refer to the weaknesses.

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

The paper proposes A-MFO-Net, a GNN-FNO-GNN framework for long-horizon 3D thermal-fluid simulation on unstructured meshes. The method uses GNNs to extract and reconstruct local features, an FNO to capture global information, and a Node-Edge Attention mechanism to weight neighbor messages. Experimental results show improvements in velocity and pressure prediction over several baselines. However, the paper has several weaknesses in methodological novelty and experimental design.

Critical strengths

The paper addresses a practical problem in 3D thermal-fluid simulation. The experiments cover 400 autoregressive prediction steps on a mesh with approximately 22,000 cells. The proposed model achieves competitive results, particularly in velocity and pressure prediction, and the authors report results across three random seeds.

Critical weaknesses

First, the proposed GNN-FNO-GNN framework mainly combines existing modules. The GNN encoder extracts local features, the FNO captures global information, and the GNN decoder reconstructs the output fields. Similar ideas have already been explored in graph-Fourier architectures such as GINO. The paper does not introduce a clear new mechanism for combining local and global information.

Second, the Node-Edge Attention mechanism is similar to existing graph attention methods, where node and edge features are used to compute attention weights. Although the authors compare against a version without attention, this comparison also changes message normalization. Comparisons with other common attention or aggregation methods are needed to show why this design is useful.

Third, the ablation experiments do not properly control variables. The FNO-only baseline removes both GNN modules while also changing the input projection, output decoder, and parameter count. Therefore, the results cannot show which changes lead to the performance improvement. The authors should separately examine the contributions of the GNN encoder, FNO, and GNN decoder.

Questions for authors

Could the authors clarify the main difference between A-MFO-Net and existing graph-Fourier architectures such as GINO?

Can the authors compare the proposed Node-Edge Attention with standard graph attention or simpler message aggregation methods under the same normalization?

Can the authors provide separate ablations for the GNN encoder, FNO, and GNN decoder while keeping other settings as consistent as possible?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.