arxiv: 2604.13369 · v1 · submitted 2026-04-15 · ⚛️ physics.comp-ph · cs.LG· physics.flu-dyn

Recognition: unknown

AeTHERON: Autoregressive Topology-aware Heterogeneous Graph Operator Network for Fluid-Structure Interaction

Sushrut Kumar

Authors on Pith no claims yet

Pith reviewed 2026-05-10 12:36 UTC · model grok-4.3

classification ⚛️ physics.comp-ph cs.LGphysics.flu-dyn

keywords fluid-structure interactiongraph neural operatorimmersed boundary methodautoregressive extrapolationsurrogate modelingvortex topologyheterogeneous graphflexible membrane

0 comments

The pith

AeTHERON mirrors immersed boundary stencils in a dual-graph operator to extrapolate fluid-structure vortex flows with mean error 0.168.

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

The paper develops AeTHERON to create fast surrogate models for flows where flexible bodies drive chaotic fluid motion and vice versa. It builds the network architecture around the physical structure of immersed boundary methods by using two separate graphs for fluid and structure, linked by sparse attention that copies the local support of physical interpolation rules. Sinusoidal embeddings handle time so the model can continue predictions past its short training window. A reader would care because full simulations of these coupled problems take hours per case while this approach runs in milliseconds, opening the way to scan many design parameters quickly.

Core claim

AeTHERON is a heterogeneous graph neural operator whose dual-graph representation separates fluid and structural domains and couples them through sparse cross-attention that directly reflects the compact support of immersed boundary method stencils, together with sinusoidal time embeddings, so that training on the first 150 timesteps of a flapping flexible caudal fin simulation allows autoregressive extrapolation to later times with a mean absolute error of 0.168 while preserving large-scale vortex topology and wake structure across a 4x5 grid of membrane thickness and Strouhal number.

What carries the argument

Dual-graph representation of fluid and structure coupled by sparse cross-attention that copies immersed boundary method stencil support, placed inside a shared latent space with continuous time embeddings.

If this is right

The model produces millisecond-per-timestep inference on GPU while matching the qualitative vortex and wake features of full direct numerical simulation.
Error remains bounded at 0.168 mean absolute value when extrapolating 50 timesteps beyond the training window without any retraining.
Prediction error peaks at flapping half-cycle transitions, matching the physical times of fastest flow reorganization.
Continuous sinusoidal time embeddings allow the same weights to handle arbitrary lead times after the initial 150 steps.
The same trained weights work across the full 4x5 grid of membrane thickness and Strouhal number values.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same stencil-mirroring cross-attention pattern could be reused for other multiphysics problems that involve immersed moving interfaces.
Rapid parameter sweeps in bio-inspired propulsion design become feasible once a single short training segment is available.
Hybrid schemes could combine this operator with occasional full simulations at the high-error transition points to maintain long-term accuracy.
The approach suggests that embedding known numerical stencil geometry into graph attention layers may stabilize autoregressive predictions in other chaotic flow regimes.

Load-bearing premise

The dual-graph architecture with IBM-mirroring sparse cross-attention and sinusoidal time embeddings supplies enough inductive bias for reliable autoregressive rollout in chaotic unsteady flows past the short training window.

What would settle it

Running the trained model on a new flapping frequency or membrane thickness outside the 4x5 grid and observing rapid growth in wake error or loss of vortex topology within 20 timesteps would show the inductive bias is insufficient.

Figures

Figures reproduced from arXiv: 2604.13369 by Sushrut Kumar.

**Figure 1.** Figure 1: 2.1 Encoder The input features of the membrane and the flow can lie in vector spaces of different dimensions. Therefore, we design an encoder stage to project the input features from the physical domains into a shared high-dimensional latent space. This enables us to model the complex nonlinear interactions between fluid flow and membrane motion within a shared space. Moreover, this lifting enriches the 3 … view at source ↗

**Figure 1.** Figure 1: AeTHERON problem formulation and physical motivation. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Figure 2: Full simulation timeline showing the train/validation split and extrapolation performance of [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Qualitative comparison of AeTHERON predictions (orange) against ground truth DNS [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Schematics of computational setup. (a) Rectilinear mesh of the flow domain, (b) Un [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

**Figure 5.** Figure 5: Isosurface of Q at Q = 5 colored by ωzC/U∞ for three representative cases: highly flexible (Fin-HF, h ∗ = 0.02), intermediately flexible (Fin-IF, h ∗ = 0.04), and rigid (Fin-R). Top and bottom rows show two viewing angles. The flapping cycle progresses left to right, illustrating leading-edge vortex formation, mid-stroke shedding, and wake evolution [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

**Figure 6.** Figure 6: Hydrodynamic performance across fin flexibility and Strouhal number. (a) Thrust coefficient [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

read the original abstract

Surrogate modeling of body-driven fluid flows where immersed moving boundaries couple structural dynamics to chaotic, unsteady fluid phenomena remains a fundamental challenge for both computational physics and machine learning. We present AeTHERON, a heterogeneous graph neural operator whose architecture directly mirrors the structure of the sharp-interface immersed boundary method (IBM): a dual-graph representation separating fluid and structural domains, coupled through sparse cross-attention that reflects the compact support of IBM interpolation stencils. This physics-informed inductive bias enables AeTHERON to learn nonlinear fluid-structure coupling in a shared high-dimensional latent space, with continuous sinusoidal time embeddings providing temporal generalization across lead times. We evaluate AeTHERON on direct numerical simulations of a flapping flexible caudal fin, a canonical FSI benchmark featuring leading-edge vortex formation, large membrane deformation, and chaotic wake shedding across a 4x5 parameter grid of membrane thickness (h* = 0.01-0.04) and Strouhal number (St = 0.30-0.50). As a proof-of-concept, we train on the first 150 timesteps of a representative case using a 70/30 train/validation split and evaluate on the fully unseen extrapolation window t=150-200. AeTHERON captures large-scale vortex topology and wake structure with qualitative fidelity, achieving a mean extrapolation MAE of 0.168 without retraining, with error peaking near flapping half-cycle transitions where flow reorganization is most rapid -- a physically interpretable pattern consistent with the nonlinear fluid-membrane coupling. Inference requires milliseconds per timestep on a single GPU versus hours for equivalent DNS computation. This is a continuously developing preprint; results and figures will be updated in subsequent versions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

AeTHERON has a reasonable physics-mirrored graph architecture for FSI but only shows quantitative results on one case despite claiming a full parameter grid.

read the letter

The core idea here is a heterogeneous graph operator that splits fluid and structure into separate graphs and uses sparse cross-attention to echo the local stencils of the immersed boundary method, plus sinusoidal embeddings for time. That design choice gives a clear inductive bias for coupled problems and is the main novelty. On the caudal fin example it produces qualitatively plausible vortex and wake patterns in autoregressive rollout, and the reported inference speed is a practical advantage over DNS. Those pieces are worth noting for anyone building surrogates in this area. The evaluation, however, stays narrow. The abstract references a 4x5 grid of membrane thickness and Strouhal number, yet the mean extrapolation MAE of 0.168 and the qualitative description both come from a single representative case trained on the first 150 timesteps. No per-parameter numbers, no aggregate statistics across the grid, and no baseline comparisons appear. In unsteady FSI the wake and vortex behavior can change noticeably with those parameters, so one rollout does not establish that the architecture generalizes. The note that error rises near half-cycle transitions is physically sensible but does not fix the scope problem. This work is aimed at people who already follow graph neural operators for fluids and want to see how IBM structure can be baked in. A reader hunting for new architectural motifs might pull a useful idea or two, but anyone needing validated performance across conditions will find the evidence too thin. I would send it to peer review if the authors add the missing multi-parameter results and at least one baseline; right now the central claim outruns what is shown.

Referee Report

3 major / 1 minor

Summary. The paper introduces AeTHERON, a heterogeneous graph neural operator for surrogate modeling of fluid-structure interaction in flapping flexible caudal fin flows. Its architecture uses a dual-graph representation (fluid and structure domains) coupled by sparse cross-attention that mirrors immersed-boundary stencils, plus sinusoidal time embeddings for temporal generalization. Trained on the first 150 timesteps of one representative DNS case (70/30 split), it is evaluated autoregressively on the unseen window t=150-200 and claims to capture large-scale vortex topology with qualitative fidelity while achieving a mean extrapolation MAE of 0.168 without retraining across the full 4x5 (h*, St) benchmark grid.

Significance. A validated architecture that supplies sufficient inductive bias for stable, long-horizon extrapolation in chaotic unsteady FSI would be a useful contribution to physics-informed surrogate modeling, especially given the reported millisecond-per-timestep inference speed relative to DNS. The explicit mapping of the dual-graph and cross-attention to IBM stencils is a clear strength if it demonstrably improves generalization.

major comments (3)

[Abstract] Abstract: the central claim that AeTHERON achieves the reported performance 'across a 4x5 parameter grid' without retraining is unsupported. The methods description states that training (70/30 split on first 150 timesteps) and evaluation (t=150-200 rollout) were performed exclusively on 'a representative case'; no per-parameter or aggregate quantitative results are shown for the remaining 19 combinations of h* and St.
[Abstract] Abstract and Results: the single reported mean extrapolation MAE of 0.168 is given only for one case. In chaotic unsteady FSI, vortex topology and wake shedding vary sharply with h* and St; therefore the single-case metric does not establish that the dual-graph + sparse cross-attention + sinusoidal embedding inductive bias suffices for the full benchmark.
[Abstract] Abstract: no baseline comparisons (standard GNN operators, PINNs, or other FSI surrogates), error bars, or quantitative results across the parameter grid are provided, and details on training procedure, loss, stability, or hyper-parameters are absent, preventing assessment of whether the claimed inductive bias is responsible for the observed behavior.

minor comments (1)

[Abstract] Abstract: the statement that this is a 'continuously developing preprint; results and figures will be updated' is atypical for a submitted manuscript and should be removed or clarified.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback, which highlights important clarifications needed in this early-stage preprint. We agree that the abstract overstates the current scope of results and will revise it to accurately reflect the proof-of-concept nature of the work on a single representative case. The core architectural contributions remain as described, but we will improve transparency regarding limitations and planned extensions.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that AeTHERON achieves the reported performance 'across a 4x5 parameter grid' without retraining is unsupported. The methods description states that training (70/30 split on first 150 timesteps) and evaluation (t=150-200 rollout) were performed exclusively on 'a representative case'; no per-parameter or aggregate quantitative results are shown for the remaining 19 combinations of h* and St.

Authors: We agree that the abstract phrasing is ambiguous and can be read as claiming quantitative performance across the full grid. The 4x5 grid is the benchmark from which the DNS data were drawn, but the reported training, autoregressive evaluation, MAE of 0.168, and qualitative vortex capture are strictly for one representative case. The phrase 'without retraining' refers to temporal extrapolation on unseen timesteps within that case. We will revise the abstract to explicitly state the single-case scope and note that full-grid evaluation is planned for future versions. This is a textual clarification only. revision: yes
Referee: [Abstract] Abstract and Results: the single reported mean extrapolation MAE of 0.168 is given only for one case. In chaotic unsteady FSI, vortex topology and wake shedding vary sharply with h* and St; therefore the single-case metric does not establish that the dual-graph + sparse cross-attention + sinusoidal embedding inductive bias suffices for the full benchmark.

Authors: We concur that a single-case result cannot demonstrate parameter-space generalization, given the known sensitivity of vortex dynamics to h* and St. The manuscript presents these results as an initial proof-of-concept for stable long-horizon rollout in a representative chaotic FSI problem, with error patterns that align with physical expectations (peaking at half-cycle transitions). We will revise the abstract and results to clearly delimit the scope and state that quantitative multi-parameter validation is ongoing work. We do not claim the inductive bias has been proven sufficient for the full grid at this stage. revision: yes
Referee: [Abstract] Abstract: no baseline comparisons (standard GNN operators, PINNs, or other FSI surrogates), error bars, or quantitative results across the parameter grid are provided, and details on training procedure, loss, stability, or hyper-parameters are absent, preventing assessment of whether the claimed inductive bias is responsible for the observed behavior.

Authors: The full manuscript contains a methods section with training details (70/30 split, MAE loss on velocity/pressure fields, sinusoidal embeddings, autoregressive rollout procedure, and stability observations). However, we agree these should be summarized more prominently. We will add a concise training description to the abstract and include error bars or run-to-run variance for the reported case. Baseline comparisons and full-grid quantitative results are not present in the current version; we will incorporate them in the revision where feasible through additional experiments. This addresses the assessment concern without altering the presented proof-of-concept. revision: partial

Circularity Check

0 steps flagged

No circularity: model trained on external DNS data with held-out temporal extrapolation

full rationale

The paper trains AeTHERON on DNS simulation data using a 70/30 split of the first 150 timesteps and evaluates autoregressive rollout on unseen later timesteps (t=150-200). No equations, parameters, or self-citations reduce the reported MAE or vortex-capture claim to a fitted input by construction. The architecture (dual-graph IBM mirroring, sparse cross-attention, sinusoidal embeddings) is presented as an inductive bias choice rather than a tautological redefinition of the target quantities. The single-case proof-of-concept evaluation does not collapse into the training data itself, and no load-bearing uniqueness theorem or ansatz is imported from prior self-work. This is a standard data-driven surrogate setup with external ground truth.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Without full text, specific training hyperparameters, loss weights, or architecture details cannot be audited. The central claim rests on the effectiveness of the proposed IBM-mirroring inductive bias in enabling extrapolation.

pith-pipeline@v0.9.0 · 5609 in / 1265 out tokens · 37540 ms · 2026-05-10T12:36:26.625644+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

[1]

Neural operator: Graph kernel network for partial differential equations

Anima Anandkumar, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Nikola Kovachki, Zongyi Li, Burigede Liu, and Andrew Stuart. Neural operator: Graph kernel network for partial differential equations. InICLR 2020 workshop on integration of deep neural models and differential equations,

2020
[2]

Multiscale MeshGraphNets.arXiv preprint arXiv:2210.00612,

Meire Fortunato, Tobias Pfaff, Peter Wirnsberger, Alexander Pritzel, and Peter Battaglia. Multiscale meshgraphnets.arXiv preprint arXiv:2210.00612,

work page arXiv
[3]

Deep learning-based surrogate models for parametrized pdes: Handling geometric variability through graph neural networks.arXiv preprint arXiv:2308.01602,

Nicola Rares Franco, Stefania Fresca, Federico Tombari, and Andrea Manzoni. Deep learning-based surrogate models for parametrized pdes: Handling geometric variability through graph neural networks.arXiv preprint arXiv:2308.01602,

work page arXiv
[4]

Data-driven modeling of fluid flow around rotating structures with graph neural networks.arXiv preprint arXiv:2503.22252,

Rui Gao, Zijie Cheng, and Rajeev K Jaiman. Data-driven modeling of fluid flow around rotating structures with graph neural networks.arXiv preprint arXiv:2503.22252,

work page arXiv
[5]

Flow-induced dorso-ventral deformation enhances propulsive efficiency in flexible caudal fins.Bioinspiration & Biomimetics, 21(1):016027, 2026a

Sushrut Kumar, Matthew J McHenry, Jung-Hee Seo, and Rajat Mittal. Flow-induced dorso-ventral deformation enhances propulsive efficiency in flexible caudal fins.Bioinspiration & Biomimetics, 21(1):016027, 2026a. Sushrut Kumar, Joshua Romero, Jung-Hee Seo, Massimiliano Fatica, and Rajat Mittal. A gpu- accelerated sharp interface immersed boundary solver for...

2026
[6]

Freeman scholar lecture (2021)—sharp-interface immersed boundary methods in fluid dynamics.Journal of Fluids Engineering, 147(3):030801,

Rajat Mittal, Jung-Hee Seo, Jacob Turner, Sushrut Kumar, Suryansh Prakhar, and Ji Zhou. Freeman scholar lecture (2021)—sharp-interface immersed boundary methods in fluid dynamics.Journal of Fluids Engineering, 147(3):030801,

2021
[7]

Rigno: A graph-based framework for robust and accurate operator learning for pdes on arbitrary domains.arXiv preprint arXiv:2501.19205,

Sepehr Mousavi, Shizheng Wen, Levi Lingsch, Maximilian Herde, Bogdan Raoni´c, and Siddhartha Mishra. Rigno: A graph-based framework for robust and accurate operator learning for pdes on arbitrary domains.arXiv preprint arXiv:2501.19205,

work page arXiv
[8]

A mixture of experts gating network for enhanced surrogate modeling in external aerodynamics.arXiv preprint arXiv:2508.21249,

Mohammad Amin Nabian and Sanjay Choudhry. A mixture of experts gating network for enhanced surrogate modeling in external aerodynamics.arXiv preprint arXiv:2508.21249,

work page arXiv
[9]

Data-efficient time-dependent pde surrogates: Graph neural simulators vs

Dibyajyoti Nayak, Qingwen Zhang, Tamer A Yousif, Liang Zhang, and Peyman Givi. Data-efficient time-dependent pde surrogates: Graph neural simulators vs. neural operators.arXiv preprint arXiv:2509.06154,

work page arXiv
[10]

State space neural operator.arXiv preprint arXiv:2507.23428,

Levi Nguyen, Matthew Wellborn, and Benjamin Peherstorfer. State space neural operator.arXiv preprint arXiv:2507.23428,

work page arXiv
[11]

DoMINO: A Decomposable Multi-scale Iterative Neural Operator for Modeling Large Scale Engineering Simulations,

Rishikesh Ranade, Mohammad Amin Nabian, Kaustubh Tangsali, Alexey Kamenev, Oliver Hennigh, Ram Cherukuri, and Sanjay Choudhry. Domino: A decomposable multi-scale iterative neural operator for modeling large scale engineering simulations.arXiv preprint arXiv:2501.13350,

work page arXiv
[12]

Gpu accelerated digital twins of the human heart open new routes for cardiovascular research.Scientific reports, 13 (1):8230, 2023a

Francesco Viola, Giulio Del Corso, Ruggero De Paulis, and Roberto Verzicco. Gpu accelerated digital twins of the human heart open new routes for cardiovascular research.Scientific reports, 13 (1):8230, 2023a. Francesco Viola, Giulio Del Corso, and Roberto Verzicco. High-fidelity model of the human heart: an immersed boundary implementation.Physical Review...

2022
[13]

Equivariant graph neural operator for modeling 3d dynamics.arXiv preprint arXiv:2401.11037,

Minkai Xu, Jiaqi Han, Aaron Lou, Jean Kossaifi, Arvind Ramanathan, Kamyar Azizzadenesheli, Jure Leskovec, Stefano Ermon, and Anima Anandkumar. Equivariant graph neural operator for modeling 3d dynamics.arXiv preprint arXiv:2401.11037,

work page arXiv
[14]

URLhttps://doi.org/10.1088/1748-3190/ad3a4e

doi: 10.1088/ 1748-3190/ad3a4e. URLhttps://doi.org/10.1088/1748-3190/ad3a4e. Ji Zhou, Jung-Hee Seo, and Rajat Mittal. Effect of hydrodynamic wakes in dynamical models of large-scale fish schools.Physics of Fluids, 37(1), 2025a. Ji Zhou, Jung-Hee Seo, and Rajat Mittal. Hydrodynamically beneficial school configurations in carangiform swimmers: insights from...

work page doi:10.1088/1748-3190/ad3a4e 2025