Pith. sign in

REVIEW 5 major objections 9 minor 44 references

FIGNN: Feature-Specific Interpretability for Graph Neural Network Surrogate Models

T0 review · 5 major / 9 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FIGNN adds a per-variable spatial mask to any message-passing GNN surrogate.

desk verdict Per-feature masks are a real extension, but the 'add-on' claim is false: FIGNN changes the prediction, so masks explain the new model, not the frozen baseline. read the letter →

arxiv 2506.11398 v1 pith:4CZVSHLD submitted 2025-06-13 cs.LG physics.flu-dyn

Riddhiman Raut , Romit Maulik , Shivam Barwey This is my paper
classification cs.LGphysics.flu-dyn
keywords graphneuralnetworksinterpretabilityTop-Kpoolingsurrogatemodelingfeature-specificmaskserrortaggingclimatedynamicsbackward-facingstepflow
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

FIGNN is an add-on for graph neural network surrogates that predict several physical fields at once. It attaches one differentiable Top-K pooling branch to each predicted feature, so each branch learns a binary spatial mask that marks the nodes most responsible for that feature's forecast error. The masks are trained by a budget term that rewards concentrating each feature's mean-squared error inside its own mask, while the frozen baseline keeps producing the forecast. Tested on an atmospheric circulation model and a backward-facing step flow, the method produces masks that trace physically recognizable structures such as frontal boundaries, jet bands, the separation bubble, and shed vortices, with predictive accuracy at or near the baseline. The paper's claim is that per-feature interpretability can be built into the forward pass instead of added after training, without giving up forecast skill.

What carries the argument

The load-bearing mechanism is the feature-specific Top-K pooling branch. For each feature $f$, a score $s_{if} = \sigma(w_f^\top h_i)$ ranks nodes from the frozen embedding $h_i$; the top $K\%$ of nodes define a feature-specific subgraph and mask $m_f$, which a down-up multiscale message-passing stack refines and the frozen decoder turns into that feature's output channel. The training objective couples the masks to error through $L = \text{MSE} + \lambda \sum_f 1/\text{Budget}_f$ with $\text{Budget}_f = \text{MSE}(m_f \odot x_{\text{pred},f}, m_f \odot x_{\text{target},f})$, so masks are rewarded for containing their own feature's forecast error. The architecture is agnostic to the baseline surrogate, provided it follows the encode-process-decode pattern.

What would settle it

On any held-out snapshot, compute the per-feature budget for the learned mask and compare it with the average budget of many random masks of identical size; if the learned mask's error fraction is not reliably higher, the claim that masks localize per-feature forecast error is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a single global importance mask entangles the spatial signatures of different output variables, and that replacing it with a feature-specific masking module is enough to disentangle them. The feature-specific interpretability module gives each output variable its own Top-K pooling branch, scored by a learned projection of the frozen encoder-processor's node embeddings; after a down-and-up message-passing pass, the frozen decoder maps each branch back to one channel, and the full prediction is the concatenation of the feature-wise outputs. A budget regularizer, $1/\text{Budget}_f$, is added to the MSE loss, where $\text{Budget}_f$ measures the error inside the mask for feature $f$. The reported result is that the masks concentrate each feature's own error while remaining human-readable: on the atmospheric data they organize into bands along jet and convection zones, and on the separated flow they stay anchored to the recirculation bubble, with $u_x$ and $u_y$ masks tracking different aspects of the same shear-layer physics.

Load-bearing premise

The central premise is that the frozen surrogate's shared node embeddings already contain enough information for separate learned projections, after extra message passing, to disentangle each variable's spatial influence; if those embeddings lack that information, the per-feature masks are arbitrary re-weightings rather than true attributions.

Editorial extensions

If this is right

  • A single forward pass through a frozen surrogate equipped with FIGNN returns both the multivariate forecast and a stack of feature-specific masks, so interpretability adds no inference-time cost.
  • Per-feature masks turn error tagging into a variable-by-variable diagnostic: temperature errors can be localized separately from humidity or wind errors, enabling feature-aware sensor placement or mesh refinement.
  • Masks can be tuned by the budget coefficient $\lambda$ and the node reduction factor RF: larger $\lambda$ captures more error inside the mask but fragments it spatially, while moderate $\lambda$ trades a little error coverage for physically contiguous regions.
  • On the two demonstrated systems, masks remain coherent and stable across node budgets, concentrating on the regions where prediction errors actually concentrate, such as the separation bubble behind the step.
  • The framework is format-agnostic: the same recipe works on a structured latitude-longitude climate grid and on an unstructured finite-volume mesh.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors leave implicit a natural downstream use: the per-feature budget scores could drive feature-aware adaptive mesh refinement, refining cells where a single variable's mask concentrates error rather than refining on total error.
  • The same frozen-embedding design could in principle attach to any encoder-processor-decoder surrogate, including attention-based graph transformers; testing whether masks remain coherent under a different backbone would check the architecture-agnostic claim.
  • A quantitative test of physical meaningfulness is straightforward but not performed in the paper: compare mask regions to independently computed physical markers, such as the reattachment line in BFS or the jet axis in SPEEDY, and measure their overlap.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 9 minor

Summary. FIGNN augments a frozen, pre-trained message-passing GNN surrogate with N_F parallel feature-wise Top-K pooling branches. Each branch computes node importance scores from the frozen embeddings, pools a feature-specific sub-graph, refines it through down/up multiscale message-passing stacks, and uses the frozen decoder to produce that feature's output channel; the full prediction is assembled channel-wise from the branch outputs. The training objective adds a budget-regularization term, Eq. (7), that penalizes the inverse of the per-feature masked MSE defined in Eq. (8). The method is evaluated on two datasets: the SPEEDY atmospheric emulator with four output variables, and a backward-facing-step (BFS) OpenFOAM flow with two velocity components. Evidence consists of training loss curves, visual rollout snapshots, mask overlays, and budget curves as a function of the regularization coefficient and node reduction factor.

Significance. If validated, the feature-specific masking idea is a useful generalization of the single-mask error-budget framework of prior work, and the two physically distinct benchmark domains support the generality claim. The paper contributes a clean architectural template, a clear extension of budget regularization to multiple masks, and complementary visualization of masks on both a structured climate grid and an unstructured CFD mesh. However, the current evidence is mostly qualitative: there is no held-out accuracy table, no repeated-seed statistics, and the budget metric is essentially the quantity being optimized. The central claim that FIGNN is an interpretability add-on preserving the frozen baseline is also not supported by the architecture, which changes the prediction. These issues are load-bearing for the abstract's promises, so the manuscript needs substantive revision before the contribution can be assessed as stated.

major comments (5)
  1. [§1 and §2, Eqs. (3)–(6)] The claim that FIGNN is an interpretability add-on that 'keeps runtime identical to the frozen baseline' is not supported by the architecture. The final prediction is assembled from branch-modified embeddings: x'_{i,f} = (D(h^{up}_f))_{i,f} with h^{up}_f = M^{up}_f(U_f(M^{down}_f(pool_f(h))) + h), so in general x'_{i,f} differs from the frozen-baseline output (D(h))_{i,f}. The masks therefore explain the augmented FIGNN model, not the frozen baseline, and inference runs N_F branch processors in addition to the baseline forward pass. This mismatch between the stated contribution and the delivered object should be resolved either by changing the architecture so the baseline output is preserved (e.g., by using masks as gating on baseline predictions) or by reframing the contribution as a new jointly trained surrogate with built-in feature attribution and reporting the true inference cost.
  2. [§2, Eqs. (7)–(8); §3.1, Figs. 7–9; §3.2, Fig. 12] The budget metric is essentially the training objective itself. Budget_f is defined as the MSE restricted to the masked nodes, and the loss adds λ Σ_f 1/Budget_f; hence larger λ directly forces larger masked-node MSE. Reporting that higher λ yields a higher budget and more fragmented masks is a direct consequence of the loss, not an independent finding. The paper should validate error localization with metrics that are not optimized, such as the error coverage of the masks under the frozen baseline, comparison against random or degree-based mask selection, or a faithfulness measure computed on held-out data after training.
  3. [§3, Figs. 4, 5, 10] The abstract claims 'competitive predictive performance,' but no quantitative test accuracy is reported. Training loss curves and visual rollouts are insufficient to compare against the baseline, especially since the FIGNN models have substantially more parameters (14.5M vs 1.97M on SPEEDY and 7.27M vs 1.97M on BFS, §2.1). Please provide held-out test-set error metrics (e.g., per-feature RMSE or MAE, and rollout-step-wise errors) with repeated-seed statistics for each λ and RF setting.
  4. [§2, Eq. (2)] The paper does not specify how gradients are computed through the non-differentiable Top-K selection and the binary mask m_f used in Eq. (8). Because the mask is discrete and the budget term depends on it, a straight-through estimator, Gumbel relaxation, or another mechanism must be described for reproducibility. Please state the exact training-time treatment of the mask and whether the same binarization is used at inference.
  5. [§3.1 and §3.2, Figs. 6, 7, 11] The interpretability validation is entirely qualitative. Statements that the masks align with 'frontal boundaries,' 'jet cores,' 'separation bubbles,' or 'vortex cores' are visual assessments. Without a quantitative coherence metric or a comparison to physics-based importance fields (e.g., error magnitude, gradient strength, vorticity), the central claim that the masks are physically meaningful is not fully supported.
minor comments (9)
  1. [§1] The opening contribution bullet contains a typo: 'They key contributions' should read 'The key contributions'.
  2. [§2, after Eq. (2)] The sentence defining h_i reads 'where h_i∈R^128 is the is the node embedding'; the duplicated phrase should be removed.
  3. [§2, Eq. (2) and Fig. 2] The notation is inconsistent: Eq. (2) uses a learnable vector w_f, while Fig. 2(a) uses a projection vector p and writes scores as σ(X·p/∥p∥). Please unify the notation and clarify the normalization.
  4. [§2.1] The phrase 'underlining graph connectivity' should read 'underlying graph connectivity'.
  5. [§3.1, Fig. 6 caption] The caption lists features as 'T,p,u_500,v_500,' but the dataset features are T, q, u_500, and v_500; the 'p' appears to be a typo for 'q'.
  6. [§2, Eq. (7)] The term 1/Budget_f is undefined if Budget_f is zero; please specify a numerical floor or epsilon in the denominator.
  7. [References] References [18] and [38] appear to be the same paper (Barwey et al., Journal of Computational Physics 495, 112537); please remove the duplicate or cite different works.
  8. [§2.1 and §3.2] The reduction factor RF is used but never formally defined; please state explicitly that RF=16 means retaining |V|/16 nodes, and specify whether the same convention is used for both datasets.
  9. [§3, Fig. 4] The caption does not explain which solid curve corresponds to which λ value or how the dashed baseline is distinguished; please add a legend or explicit caption description.

Circularity Check

1 steps flagged · score 5.0 of 10

The error-budget validation is definitional: Eq. (8)'s Budget_f is exactly the quantity minimized through 1/Budget_f in Eq. (7), so the reported 'higher λ ⇒ higher budget' trend restates the training objective rather than independently confirming error localization.

  1. self definitional [Section 2, Eqs. (7)-(8); Section 3.1 Figure 8 and Section 3.2 Figure 12]
    "the loss function is augmented with a regularization term that explicitly ties the importance masks to prediction error. The total loss is defined as: L=MSE(x pred,x target)+λ∑_{f=1}^{N_F} 1/Budget_f ... Budget_f = MSE(m_f⊙x pred,f,m_f⊙x target,f). ... Together, these budget analyses reveal a fundamental trade-off: higher regularization λ enhances interpretability by forcing the model to isolate a greater fraction of the error within feature-specific masks."

    The budget metric is not an independent diagnostic: Eq. (8) defines Budget_f as the masked MSE on the selected nodes, and Eq. (7) penalizes 1/Budget_f for each feature. Minimizing Eq. (7) with λ>0 directly forces Budget_f upward (subject to the fixed K% cardinality), so observing larger budgets at larger λ merely confirms that the optimizer minimized the term it was given. The paper's stated contribution that the budget term 'guides each feature-mask to contain and localize its own variable-specific forecast error' is therefore true by construction, not by empirical demonstration.

full rationale

The one genuinely circular element is the error-budget validation: Eq. (8)'s Budget_f is exactly the masked-MSE quantity whose reciprocal is penalized in Eq. (7), so training with λ>0 is precisely an optimization of the budget. Reporting that larger λ produces larger budgets (Figs. 8, 9, 12) restates the objective rather than independently confirming that masks localize error. The feature-specific mask patterns, rollout stability, and predictive-accuracy comparisons are independent content and not circular. The claim that the frozen baseline is preserved 'with runtime identical' is contradicted by Eq. (6) — the forward pass runs N_F branch processors and assembles a new prediction from modified embeddings — but that is a correctness/scope mismatch, not a circular derivation. No load-bearing self-citation or imported uniqueness theorem was found; refs. [38] and [40] are background/method reuse and do not by themselves force the paper's conclusions. Because the central interpretability and accuracy claims retain independent content, the score is kept below the 'central claim reduces to construction' threshold.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

FIGNN introduces an architectural module (FSIM), not a new postulated physical entity. There are no new particles, forces, dimensions, or conserved quantities, so the invented-entities ledger is empty. The load-bearing auxiliary items are instead the tuned hyperparameters and the domain assumptions about embedding richness and interpretability.

free parameters (4)
  • lambda (error-budget regularization coefficient) = swept 0, 1e-4, 1e-3, 1e-2, 1e-1 with no formal selection criterion
    Chosen by hand to balance budget and mask coherence; the paper's budget results depend directly on this choice.
  • reduction factor RF = 16, 8, 4
    Controls mask cardinality; selected for demonstration, not derived from any principle.
  • number of nearest neighbors k for SPEEDY graph edges = 10
    Edge construction choice stated in Sec 2.1; affects graph connectivity and therefore mask geometry.
  • message-passing stack depths L_down and L_up = not stated
    Architecture hyperparameters are not reported, hampering replication and making the parameter cost hard to audit.
assumptions (4)
  • domain assumption Frozen baseline embeddings h_i are rich enough that linear projections w_f^T h_i can separate feature-specific spatial importance.
    Eq (2) uses only frozen embeddings; if features are entangled in h_i, the per-feature masks cannot truly disentangle them.
  • ad hoc to paper Top-K binary masks can be optimized despite non-differentiable selection; no gradient surrogate is specified.
    Eqs (2)-(4) do not describe how gradients pass through the TopK operator, yet the method assumes a workable path exists.
  • domain assumption Fixed graph and edge attributes make learned spatial masks directly interpretable as physical regions.
    Sec 2 states fixed relationships enable direct interpretation, but this ignores representation ambiguity in learned embeddings.
  • domain assumption SPEEDY and BFS are representative testbeds for general scientific surrogate interpretability.
    The generality claim in the abstract rests on only two datasets with different but limited physical complexity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FIGNN: Feature-Specific Interpretability for Graph Neural Network Surrogate Models." pith.science (2026). https://pith.science/paper/4CZVSHLD

@misc{pith2026250611398,
  author       = {Pith},
  title        = {Pith review of: FIGNN: Feature-Specific Interpretability for Graph Neural Network Surrogate Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4CZVSHLD}},
  note         = {Machine review of arXiv:2506.11398}
}
read the original abstract

This work presents a novel graph neural network (GNN) architecture, the Feature-specific Interpretable Graph Neural Network (FIGNN), designed to enhance the interpretability of deep learning surrogate models defined on unstructured grids in scientific applications. Traditional GNNs often obscure the distinct spatial influences of different features in multivariate prediction tasks. FIGNN addresses this limitation by introducing a feature-specific pooling strategy, which enables independent attribution of spatial importance for each predicted variable. Additionally, a mask-based regularization term is incorporated into the training objective to explicitly encourage alignment between interpretability and predictive error, promoting localized attribution of model performance. The method is evaluated for surrogate modeling of two physically distinct systems: the SPEEDY atmospheric circulation model and the backward-facing step (BFS) fluid dynamics benchmark. Results demonstrate that FIGNN achieves competitive predictive performance while revealing physically meaningful spatial patterns unique to each feature. Analysis of rollout stability, feature-wise error budgets, and spatial mask overlays confirm the utility of FIGNN as a general-purpose framework for interpretable surrogate modeling in complex physical domains.

Figures

Figures reproduced from arXiv: 2506.11398 by the authors.

Figure 1
Figure 1. Schematic of the Feature-specific Interpretable Graph Neural Network (FIGNN). The frozen baseline model (blue) consists of encoding, [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Feature-specific Top–K pooling and mask generation. (a) One-dimensional projection of the node–feature matrix X ∈ R |V|×NF onto a learnable vector p produces scores y = σ(X · p/∥p∥). Scores are sorted and truncated, keeping only the highest-ranking fraction, which defines the set of selected (masked) nodes (yellow). (b) In FIGNN, each input feature f ∈ {1, . . . , NF} has its own projection vector pf , resulting in … view at source ↗
Figure 3
Figure 3. Feature-specific Top-K masks on the two benchmark datasets. (a) Temperature field T from the SPEEDY dataset at time t = t0. (b) Temperature mask produced by FIGNN at the same instant; yellow pixels mark nodes retained by the Top-K operator. (c) Enlargement of the red rectangle in (b), revealing the regular latitude–longitude grid and the edges that connect neighbouring nodes. (d) Stream-wise velocity Ux for a snapsh… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Training loss curves for interpretable (solid lines) versus baseline (dashed line) GNNs on the SPEEDY dataset (left) and the BFS dataset [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Multi-step rollout predictions for each atmospheric feature from the SPEEDY dataset, produced by FIGNN (RF [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Overlay of feature-specific interpretability masks on the multi-step rollouts for the SPEEDY dataset. In each panel, the mask highlights [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Effect of the error–tagging coefficient λ on feature-specific Top-K masks (RF = 16). Each row shows the fixed–size mask for four atmospheric features, while sweeping λ from 10−1 (top) to 0 (bottom).Because the reduction factor pins the mask cardinality, increasing λ on…
Figure 8
Figure 8. Figure 8: Single-step budget plots for the FIGNN model on the SPEEDY dataset. The left panel displays the total error budget (i.e., the fraction of [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Rollout budget plots for the FIGNN model on the SPEEDY dataset. This figure illustrates the evolution of the total budget over multiple [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Comparison of FIGNN predictions for the backward-facing step (BFS) dataset. In the top panels, the surrogate model’s predicted [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Full-field predictions and feature-specific Top-K masks for the backward–facing-step flow at timestep 60 (Re [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Budget plots for the backward-facing step (BFS) dataset. The top row displays the single-step budgets, indicating the fraction of the [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 32 canonical work pages

  1. [1]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, A. Stuart, K. Bhattacharya, A. Anandkumar, Multipole graph neural operator for parametric partial differential equations, Advances in Neural Information Processing Systems 33 (2020) 6755–6766

    work page 2020

  2. [2]

    L. Wu, P. Cui, J. Pei, L. Zhao, X. Guo, Graph neural networks: foundation, frontiers and applications, in: Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, 2022, pp. 4840– 4841

    work page 2022

  3. [3]

    M. M. Bronstein, J. Bruna, Y . LeCun, A. Szlam, P. Vandergheynst, Geometric deep learning: going beyond euclidean data, IEEE Signal Processing Magazine 34 (4) (2017) 18–42

    work page 2017

  4. [4]

    M. Kurz, A. Beck, B. Sanderse, Harnessing equivariance: Modeling turbulence with graph neural networks, arXiv preprint arXiv:2504.07741 (2025)

    arXiv 2025

  5. [5]

    Horie, N

    M. Horie, N. Mitsume, Physics-embedded neural networks: Graph neural pde solvers with mixed boundary conditions, Advances in Neural Information Processing Systems 35 (2022) 23218–23229

    work page 2022

  6. [6]

    X. Han, H. Gao, T. Pfaff, J.-X. Wang, L.-P. Liu, Predicting physics in mesh-reduced space with temporal atten- tion, arXiv preprint arXiv:2201.09113 (2022)

    arXiv 2022

  7. [7]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, G. E. Dahl, Neural message passing for quantum chemistry, in: International conference on machine learning, PMLR, 2017, pp. 1263–1272. 16

    work page 2017

  8. [8]

    A. Khan, M. Yamada, A. Chikane, M. Kaul, Graphmesh: Geometrically generalized mesh refinement using gnns, in: International Conference on Computational Science, Springer, 2024, pp. 120–134

    work page 2024

Show all 44 references
  1. [9]

    Sanderse, P

    B. Sanderse, P. Stinis, R. Maulik, S. E. Ahmed, Scientific machine learning for closure models in multiscale problems: A review, arXiv preprint arXiv:2403.02913 (2024)

    2024 arXiv

  2. [10]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  3. [11]

    Hamilton, Z

    W. Hamilton, Z. Ying, J. Leskovec, Inductive representation learning on large graphs, Advances in neural infor- mation processing systems 30 (2017)

    2017

  4. [12]

    Veliˇckovi´c, G

    P. Veliˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y . Bengio, Graph attention networks, arXiv preprint arXiv:1710.10903 (2017)

    2017 arXiv

  5. [13]

    P. W. Battaglia, J. B. Hamrick, V . Bapst, A. Sanchez-Gonzalez, V . Zambaldi, M. Malinowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner, et al., Relational inductive biases, deep learning, and graph networks, arXiv preprint arXiv:1806.01261 (2018)

    2018 arXiv

  6. [14]

    Sanchez-Gonzalez, J

    A. Sanchez-Gonzalez, J. Godwin, T. Pfaff, R. Ying, J. Leskovec, P. Battaglia, Learning to simulate complex physics with graph networks, in: International conference on machine learning, PMLR, 2020, pp. 8459–8468

    2020

  7. [15]

    Pfaff, M

    T. Pfaff, M. Fortunato, A. Sanchez-Gonzalez, P. Battaglia, Learning mesh-based simulation with graph networks, in: International conference on learning representations, 2020

    2020

  8. [16]

    Fortunato, T

    M. Fortunato, T. Pfaff, P. Wirnsberger, A. Pritzel, P. Battaglia, Multiscale meshgraphnets, arXiv preprint arXiv:2210.00612 (2022)

    2022 arXiv

  9. [17]

    M. Lino, S. Fotiadis, A. A. Bharath, C. D. Cantwell, Multi-scale rotation-equivariant graph neural networks for unsteady eulerian fluid dynamics, Physics of Fluids 34 (8) (2022)

    2022

  10. [19]

    Deshpande, S

    S. Deshpande, S. Bordas, J. Lengiewicz, Magnet: A graph u-net architecture for mesh-based simulations, arXiv preprint arXiv:2211.00713 (2022)

    2022 arXiv

  11. [20]

    Perera, V

    R. Perera, V . Agrawal, Multiscale graph neural networks with adaptive mesh refinement for accelerating mesh- based simulations, Computer Methods in Applied Mechanics and Engineering 429 (2024) 117152

    2024

  12. [21]

    R. Lam, A. Sanchez-Gonzalez, M. Willson, P. Wirnsberger, M. Fortunato, F. Alet, S. Ravuri, T. Ewalds, Z. Eaton- Rosen, W. Hu, et al., Learning skillful medium-range global weather forecasting, Science 382 (6677) (2023) 1416–1421

    2023

  13. [22]

    Z. Li, K. Meidani, P. Yadav, A. Barati Farimani, Graph neural networks accelerated molecular dynamics, The Journal of Chemical Physics 156 (14) (2022)

    2022

  14. [23]

    Y . Park, J. Kim, S. Hwang, S. Han, Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations, Journal of chemical theory and computation 20 (11) (2024) 4857–4868

    2024

  15. [25]

    J. Xu, A. Pradhan, K. Duraisamy, Conditionally parameterized, discretization-aware neural networks for mesh- based modeling of physical systems, Advances in Neural Information Processing Systems 34 (2021) 1634–1645

    2021

  16. [26]

    T. Li, S. Zou, X. Chang, L. Zhang, X. Deng, Predicting unsteady incompressible fluid dynamics with finite volume informed neural network, Physics of Fluids 36 (4) (2024). 17

    2024

  17. [27]

    H. Gao, M. J. Zahr, J.-X. Wang, Physics-informed graph neural galerkin networks: A unified framework for solving pde-governed forward and inverse problems, Computer Methods in Applied Mechanics and Engineering 390 (2022) 114502

    2022

  18. [28]

    R. Gao, I. K. Deo, R. K. Jaiman, A finite element-inspired hypergraph neural network: Application to fluid dynamics simulations, Journal of Computational Physics 504 (2024) 112866

    2024

  19. [29]

    Barwey, P

    S. Barwey, P. Pal, S. Patel, R. Balin, B. Lusch, V . Vishwanath, R. Maulik, R. Balakrishnan, Mesh-based super- resolution of fluid flows with multiscale graph neural networks, Computer Methods in Applied Mechanics and Engineering 443 (2025) 118072

    2025

  20. [30]

    Salehi, D

    Y . Salehi, D. Giannacopoulos, Physgnn: A physics–driven graph neural network based model for predicting soft tissue deformation in image–guided neurosurgery, Advances in Neural Information Processing Systems 35 (2022) 37282–37296

    2022

  21. [31]

    Schmidt, H

    A. Schmidt, H. Zunker, A. Heinlein, M. J. Kühn, Towards graph neural network surrogates leveraging mecha- nistic expert knowledge for pandemic response, arXiv preprint arXiv:2411.06500 (2024)

    2024

  22. [32]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-cam: visual explanations from deep networks via gradient-based localization, International journal of computer vision 128 (2020) 336–359

    2020

  23. [33]

    Z. Ying, D. Bourgeois, J. You, M. Zitnik, J. Leskovec, Gnnexplainer: Generating explanations for graph neural networks, Advances in neural information processing systems 32 (2019)

    2019

  24. [34]

    D. Luo, W. Cheng, D. Xu, W. Yu, B. Zong, H. Chen, X. Zhang, Parameterized explainer for graph neural network, Advances in neural information processing systems 33 (2020) 19620–19631

    2020

  25. [35]

    Huang, M

    Q. Huang, M. Yamada, Y . Tian, D. Singh, Y . Chang, Graphlime: Local interpretable model explanations for graph neural networks, IEEE Transactions on Knowledge and Data Engineering 35 (7) (2022) 6968–6972

    2022

  26. [36]

    Harel, C

    J. Harel, C. Koch, P. Perona, Graph-based visual saliency, Advances in neural information processing systems 19 (2006)

    2006

  27. [37]

    Zhang, P

    Y . Zhang, P. Tiˇno, A. Leonardis, K. Tang, A survey on neural network interpretability, IEEE Transactions on Emerging Topics in Computational Intelligence 5 (5) (2021) 726–742

    2021

  28. [38]

    Barwey, V

    S. Barwey, V . Shankar, V . Viswanathan, R. Maulik, Multiscale graph neural network autoencoders for inter- pretable scientific machine learning, Journal of Computational Physics 495 (2023) 112537

    2023

  29. [39]

    H. Gao, S. Ji, Graph u-nets, in: international conference on machine learning, PMLR, 2019, pp. 2083–2092

    2019

  30. [40]

    Barwey, H

    S. Barwey, H. Kim, R. Maulik, Interpretable a-posteriori error indication for graph neural network surrogate models, Computer Methods in Applied Mechanics and Engineering 433 (2025) 117509

    2025

  31. [41]

    Kucharski, F

    F. Kucharski, F. Molteni, A. Bracco, Decadal interactions between the western tropical pacific and the north atlantic oscillation, Climate dynamics 26 (2006) 79–91

    2006

  32. [42]

    R. Gao, R. K. Jaiman, Predicting fluid–structure interaction with graph neural networks, Physics of Fluids 36 (1) (2024)

    2024

  33. [43]

    Veli ˇckovi´c, R

    P. Veli ˇckovi´c, R. Ying, M. Padovano, R. Hadsell, C. Blundell, Neural execution of graph algorithms, arXiv preprint arXiv:1910.10593 (2019)

    2019 arXiv

  34. [44]

    Eliasof, E

    M. Eliasof, E. Haber, E. Treister, Pde-gcn: Novel architectures for graph neural networks motivated by partial differential equations, Advances in neural information processing systems 34 (2021) 3836–3849

    2021

  35. [45]

    S. Yun, M. Jeong, R. Kim, J. Kang, H. J. Kim, Graph transformer networks, Advances in neural information processing systems 32 (2019)

    2019

  36. [46]

    Molteni, Atmospheric simulations using a gcm with simplified physical parametrizations

    F. Molteni, Atmospheric simulations using a gcm with simplified physical parametrizations. i: Model climatol- ogy and variability in multi-decadal experiments, Climate Dynamics 20 (2003) 175–191. 18

    2003

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.