Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

GALDS: A Graph-Autoencoder-based Latent Dynamics Surrogate model to predict neurite material transport

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read GALDS claims that cargo transport through a neuron's branching tree can be predicted on unseen geometries at 1-3% mean relative error and about ten times faster inference, by learning the dynamics in a compressed latent space rather than…

desk verdict GALDS is a plausible, well-engineered surrogate for neurite transport, but the headline accuracy figures are range-normalized steady-state errors, so the dynamic accuracy claim needs more evidence. read the letter →

arxiv 2507.10871 v1 pith:ZIASFSRE submitted 2025-07-15 cs.LG cs.NAmath.NAphysics.med-ph

Tsung Yeh Hsieh , Yongjie Jessica Zhang This is my paper · ORCID
classification cs.LGcs.NAmath.NAphysics.med-ph MSC 68T0765M6092C20
keywords neuritematerialtransportsurrogatemodelgraphautoencoderNeuralODElatentspacedynamicsisogeometricanalysistrafficjam
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

This paper proposes GALDS, a surrogate model that predicts how material flows through the branching tree of a neuron without running the expensive 3D simulation. The central move is to compress the three-dimensional velocity and concentration fields into small latent vectors with graph autoencoders, then learn the time evolution of transport in that latent space with a Graph Neural ODE instead of in physical space. The authors claim this latent-space paradigm is what makes the surrogate practical: on eight geometries never seen during training it achieves mean relative error between 1.0% and 2.7% for concentration with a maximum under 5.7%, and under 3% with a maximum near 8% on asymmetric traffic-jam cases, while needing 20 times less training data, about 10 times fewer trainable parameters, six times less training time, and about tenfold faster inference than the earlier PGNN surrogate. If that holds, repeated transport queries over many neurite geometries and parameter settings become feasible at a fraction of the IGA-simulation budget.

What carries the argument

The load-bearing object is the latent-space dynamics system: a Graph Neural ODE of the form $\partial \tilde{z}/\partial t = \psi^{(z)}(\bar{z}_0, \bar{u}, c, t)$, where $\bar{z}_0$ is the concentration encoder's output at the initial condition, $\bar{u}$ is the latent velocity obtained by projecting 1D reduced-order-model results, and $c$ are distance-from-root embeddings; the predicted derivative is integrated with fourth-order Runge-Kutta, so any time step is queried by re-integration. Around this core sit four modules trained in sequence: the velocity and concentration graph autoencoders (a few thousand parameters each, with pipe and bifurcation variants sharing encoder weights to keep the latent spaces consistent), the latent projection $\psi^{(u)}$ that replaces expensive 3D velocity simulation with a cheap 1D solve, and the global graph assembly that unifies all cross-section latent vectors into one tree so prediction never re-segments the geometry. The compression claim, the data-efficiency claim, and the error-accumulation claim all ride on this same machinery.

What would settle it

Run the trained GALDS pipeline on one of the eight unseen geometries and compute per-node relative errors, each node's error divided by the local exact concentration rather than by the global max-min range, at the propagating transport front and at intermediate times before steady state. If per-node errors in low-concentration branches or steep-gradient fronts exceed roughly 15-20% while the global-normalized mean relative error stays near 2-3%, then the advertised under-8% maximum is an artifact of the normalization and the model does not resolve the transient transport it claims to predict.

Watch

Extended reading notes

Core claim

GALDS claims that neurite material transport can be predicted at usable accuracy from a small data budget if the dynamics are learned in a compressed representation. The pipeline encodes 3D velocity and concentration fields from IGA simulations into latent vectors using graph autoencoders trained separately on pipe and bifurcation segments with shared encoder parameters, maps cheap 1D reduced-order-model velocity results into the same velocity latent space through a fully connected projection $\psi^{(u)}$, and advances latent concentrations with a Graph Neural ODE of the form $\partial \tilde{z}/\partial t = \psi^{(z)}(\bar{z}_0, \bar{u}, c, t)$ integrated by fourth-order Runge-Kutta. On eight unseen neurite geometries (two zebrafish, six mouse), the full pipeline reports concentration mean relative errors of 1.0-2.7% and maxima below 5.7%, with velocity mean errors below 0.4%; in the spatially heterogeneous traffic-jam setting with asymmetric attachment rates, mean error stays under 3% with maxima between 6% and 8%. The design claims three sources of gain: compression shrinks the model (about 126k trainable parameters versus 1.5M for PGNN) and the data need (10 versus 200 IGA runs); continuous-time integration avoids the error accumulation of recurrent GNNs; and the 1D-ROM velocity input carries enough physics that the decoder only has to refine it, not infer it from scratch.

Load-bearing premise

The headline accuracy figures are global-range-normalized per Eq. 29 (error divided by the max-minus-min of the exact field) and are reported at steady-state time points in Tables 3 and 5, so the central claim assumes this metric faithfully represents local, time-resolved accuracy, and the paper's own conclusion concedes that the autoencoder smooths fine-grained local detail and that the framework assumes static geometries and laminar flow.

Editorial extensions

If this is right

  • Training a transport surrogate becomes a small-data problem: about 10 IGA runs per geometry pair replace the 200 runs the earlier PGNN needed, which is the difference between days and weeks of simulation time before the surrogate exists.
  • At inference, only a 1D centerline Navier-Stokes solve plus the skeleton connectivity are required inputs; full 3D velocity is decoded from the latent projection and concentration fields are integrated forward in the latent space, cutting evaluation from thousands of core-hours to under a minute.
  • On the eight unseen geometries, concentration mean relative error stays between 1.0% and 2.7% with maximum below 5.7%, and the residual error concentrates at bifurcations and steep-gradient fronts, the regions where the pipe-versus-bifurcation training data is imbalanced, so targeted enrichment rather than redesign is the indicated next step.
  • Because the dynamics module outputs $\partial \tilde{z}/\partial t$, any time step can be queried by re-integrating with the Runge-Kutta scheme, avoiding the error accumulation of recurrent architectures and giving a continuous trajectory rather than discrete snapshots.
  • The same latent-space recipe extends to spatially heterogeneous attachment kinetics that produce asymmetric traffic-jam profiles, by adding a parameter autoencoder and two input channels while reusing the velocity modules with no retraining.
  • If the accuracy claims transfer, repeated costly tasks such as uncertainty quantification, parameter sweeps, and inverse searches over neurite transport models become practical, because each evaluation costs seconds instead of tens of thousands of core-minutes.

Reading between the lines

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

  • The twentyfold data reduction suggests the transport fields live on a very low-dimensional manifold; comparing the learned latent coordinates against a linear basis such as proper orthogonal decomposition of the same IGA fields would test whether the autoencoder has rediscovered physical transport modes and whether the Graph Neural ODE evolves genuine dynamics on them, a comparison the paper does
  • A clean ablation would remove the 1D-ROM velocity input and feed raw geometry and parameters instead; the resulting error increase would isolate how much of GALDS' accuracy comes from the physics-informed velocity shortcut versus the autoencoder compression itself, which the paper does not report.
  • If the recipe transfers, the same encode-compress-integrate-decode pipeline should apply to any low-Reynolds branching-pipe network with steady parabolic profiles, such as vascular trees, plant roots, or microfluidic channels, because the compression relies on the pipe-plus-bifurcation template structure rather than on neuron biology.
  • The static-geometry limitation the authors state in their conclusion means GALDS predicts transport for a fixed neurite tree; using it to study development or degeneration, where the tree itself grows or retracts, would require re-encoding the geometry at each state, a natural but nontrivial extension the paper leaves open.
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

4 major / 6 minor

Summary. The paper proposes GALDS, a graph-autoencoder-based latent dynamics surrogate for 3D neurite material transport. The model compresses velocity and concentration fields into low-dimensional latent representations via separate pipe and bifurcation graph autoencoders, maps 1D reduced-order-model velocity data into the velocity latent space, and predicts concentration dynamics with a GNN-based Neural ODE in latent space. The authors train on IGA simulation data from two neuron geometries and report MRE between 1.0% and 2.7% and MaxRE below 8% on eight unseen geometries, plus results on four abnormal transport cases. They also claim a 20x smaller training dataset, roughly 10x fewer trainable parameters, about 6x faster training, and approximately 10x faster inference compared with the prior PGNN surrogate.

Significance. If the accuracy and efficiency claims hold, GALDS would be a practically useful surrogate for costly 3D neuron transport simulations, and the latent-dynamics design with 1D ROM conditioning is a sensible inductive bias. The paper has concrete strengths: the code is publicly available on GitHub, the evaluation includes eight unseen geometries and four abnormal cases, and the comparisons with PGNN explicitly list dataset size, parameter count, training time, and inference time. The authors also appropriately acknowledge limitations in Section 6, such as the autoencoder smoothing localized features and the restriction to static geometries. However, the headline accuracy claim is currently not supported by the reported metric, because Eq. (29) computes a global-range-normalized RMSE rather than a local relative error, and the reported errors are evaluated at steady-state time points rather than over full trajectories. These are load-bearing issues for the central empirical claim.

major comments (4)
  1. [Eq. (29); Tables 3 and 5] The headline accuracy metric is not a relative error. Eq. (29) divides the RMSE by the global max-min range of the exact field, so it measures range-normalized RMSE rather than per-node relative error. In low-concentration regions, a node can have a large pointwise relative error and yet contribute little to the reported MRE. The abstract's claim of "mean relative error of 3%" is therefore not directly supported by Tables 3 and 5. Please report per-node relative errors with a stated rule for near-zero reference values, and also bin the errors by exact field magnitude in both the normal transport and traffic-jam cases.
  2. [Sections 4.3-4.4, Eq. (22); Tables 3 and 5] The paper advertises a dynamic surrogate, but all reported errors are evaluated at steady-state time points, as is visible in the Figure 5-7 captions and in Tables 3 and 5. The inference path in Eqs. (26)-(28) is a recursive rollout, while the training loss in Eq. (22) is a one-step loss comparing the integrated state from the encoded true state with the next encoded true state. No experiment measures error accumulation over the full trajectory. Please report MRE(t) and MaxRE(t) over all time steps for each unseen geometry and compare recursive rollout with one-step predictions. This is also needed to support the claim in the abstract and Section 3 that the Neural ODE component "effectively mitigates the issue of error accumulation."
  3. [Sections 5.1.2 and 5.2.2] The accuracy and efficiency results are single-run point estimates. No seed variation, initialization, or dataset-split variation is reported. Since GNN training is stochastic, please report the mean and standard deviation over at least three independent training runs for the headline metrics (MRE, MaxRE, and inference time) on at least one geometry per case, or clearly state that a fixed seed was used and assess sensitivity to it.
  4. [Tables 3 and 4] The comparison against PGNN does not state whether PGNN was retrained on the same training data and evaluated on the same eight unseen geometries under the same protocol, or whether the values are taken from Ref. [48]. Without a controlled comparison, the claimed 10x inference speedup and 20x dataset reduction are not fully established. Please specify the comparison protocol, or provide a same-code, same-split baseline.
minor comments (6)
  1. [Eq. (27)] The displayed recursive definition of tilde-z(t_i) is circular: for t_i > 0 it defines tilde-z(t_i) in terms of tilde-z(t_i). It should refer to the integrated previous state, for example tilde-z(t_i) = Int(psi(...)(tilde-z(t_{i-1}), ...), Delta t).
  2. [Abstract; Section 5.1.2] The abstract states "mean relative error of 3%", but Tables 3 and 5 report maximum MRE values of 2.7%; state the exact range and make clear that the metric is the range-normalized RMSE of Eq. (29).
  3. [Figure 6] The caption calls the color maps "nodal relative error", which conflicts with the global-range normalization in Eq. (29); state which formula was used to generate these maps.
  4. [Table 5] The time columns for the pipe cases are formatted ambiguously ("6,475 557 0.33"); separate the IGA, training, and inference values so the comparison is readable.
  5. [Section 1; Section 4.3] The claim that this is the first combination of GNNs with Neural ODEs for predicting physical system dynamics is weakened by Ref. [57], which already applies graph neural ODEs to trajectory extrapolation and traffic forecasting; please soften the novelty statement.
  6. [Throughout] There are numerous typographical errors, including "neurtie", "generalizabilty", "GLADS" for GALDS in Section 5.1.3, and "refurring" in the Table 5 caption; a careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: GALDS is a supervised surrogate trained and tested against independent IGA ground truth, with self-citations serving as tools and baselines rather than load-bearing proof.

full rationale

The derivation chain is self-contained as a supervised learning pipeline. The IGA solver (Section 2) generates ground-truth velocity and concentration fields; the autoencoders (Eqs. 15-16) are trained to reconstruct those fields; the latent transformation model (Eq. 18) is trained to map 1D ROM velocities to the encoded 3D velocity latent; and the latent dynamics model (Eq. 22) is trained against encoded IGA next-time-step states. At inference (Eqs. 24-28), predictions are decoded from latent trajectories and compared to IGA ground truth on unseen geometries (Tables 3 and 5). No equation defines a predicted quantity in terms of the same quantity, and no fitted parameter is renamed as a prediction: the fitted parameters are network weights, while the predicted fields are decoded outputs evaluated against independent simulation data. Self-citations to [19], [48], and [49] supply the data generator, the PGNN baseline, and the node-extraction convention; these are tools/comparators, not evidence used to prove GALDS accuracy, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the architecture. The Graph Neural ODE and sigmoid choice cite external works [57,58]. Two textual issues are noted but are not circularity: Eq. (27) has a self-referential typo in the ti > 0 branch (it should reference the previously integrated state, as Eq. 26 and Algorithm 1 make clear), and Eq. (29) normalizes error by the global max-min range, so the reported 'MRE' is a range-normalized RMSE, with Tables 3 and 5 reporting steady-state time points and no explicit recursive rollout error. These are accuracy-reporting and clarity concerns, not instances of the derivation reducing to its inputs.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central model is a data-driven surrogate. The main things the reader pays for upstream are the governing PDE assumptions inherited from the authors' IGA framework, the template-based compression, and the physics-informed 1D ROM input. No new physical entity is postulated. The free parameters are mostly architecture and normalization choices whose values are not independently justified.

free parameters (5)
  • Latent space dimension per cross-section = velocity: 3; concentration: 1 (inferred from Fig. 4 dimensions)
    Chosen by hand; no ablation is reported. All downstream accuracy depends on this compression ratio, and the value is not explicitly stated as a tuned hyperparameter.
  • Pipe/bifurcation sampling templates = 17 nodes per pipe cross-section; 23 nodes per bifurcation
    Inherited from prior PGNN work [48]. These fixed templates assume that 17 or 23 samples capture the full cross-sectional field, with no sensitivity analysis for unseen or deformed geometries.
  • Training dataset scale and split = 2 geometries x 10 boundary conditions; 80/20 train/test split
    Hand-chosen. The data-efficiency claim is relative to this small dataset, and generalization to other physical regimes such as different Péclet numbers or time scales is untested.
  • Error normalization range = global max-min of exact solution
    Eq. (29) divides the RMS error by max||cexact|| - min||cexact||. This chosen normalization controls the reported MRE and MaxRE values and is not a local relative error.
  • Model hyperparameters
    Learning rate 1e-4, full-batch Adam, 4 GConv layers plus one feedforward layer in the autoencoders, 5 GConv layers in the dynamics model, and sigmoid/softplus activations. These were chosen via internal testing without reported grid search or seed variance.
assumptions (5)
  • domain assumption The governing PDE system in Eq. (1), with the unipolar filament assumption that ignores n-, k- and v- terms, accurately models neurite material transport.
    Section 2 states the filament system is unipolar and retrograde terms are ignored. All training data are generated from this model, so biological fidelity is assumed rather than validated against experiments.
  • domain assumption Steady incompressible Navier-Stokes with no-slip walls and a parabolic inlet profile describes the flow in neurites.
    Invoked in Section 2, Eqs. (3)-(4). The low-Reynolds laminar assumption is standard for these geometries, but the surrogate inherits it without testing higher-Reynolds or unsteady regimes.
  • ad hoc to paper Pipe and bifurcation templates with 17 and 23 sampled points per cross-section are sufficient to represent the 3D fields.
    Introduced in Section 3 and Fig. 2. The entire autoencoder operates on these templates; if small-scale cross-sectional features matter, the latent compression can miss them.
  • ad hoc to paper The 1D ROM velocity u1D plus a trained decoder is sufficient to reconstruct 3D velocity, including curvature, tapering, and branching effects.
    Section 3 and Eq. (19). This is the key inductive bias behind the low-data regime: the model deliberately does not learn velocity from raw geometry, and the claim depends on the decoder recovering secondary flow effects that 1D ROM omits.
  • ad hoc to paper Latent dynamics are smooth enough for a GNN Neural ODE with RK4 integration to advance accurately over the full time range.
    Section 4.3. No smoothness or stability analysis is provided. The claim that the Neural ODE mitigates error accumulation rests on this assumption, yet only steady-state errors are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GALDS: A Graph-Autoencoder-based Latent Dynamics Surrogate model to predict neurite material transport." pith.science (2026). https://pith.science/paper/ZIASFSRE

@misc{pith2026250710871,
  author       = {Pith},
  title        = {Pith review of: GALDS: A Graph-Autoencoder-based Latent Dynamics Surrogate model to predict neurite material transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZIASFSRE}},
  note         = {Machine review of arXiv:2507.10871}
}
read the original abstract

Neurons exhibit intricate geometries within their neurite networks, which play a crucial role in processes such as signaling and nutrient transport. Accurate simulation of material transport in the networks is essential for understanding these biological phenomena but poses significant computational challenges because of the complex tree-like structures involved. Traditional approaches are time-intensive and resource-demanding, yet the inherent properties of neuron trees, which consists primarily of pipes with steady-state parabolic velocity profiles and bifurcations, provide opportunities for computational optimization. To address these challenges, we propose a Graph-Autoencoder-based Latent Dynamics Surrogate (GALDS) model, which is specifically designed to streamline the simulation of material transport in neural trees. GALDS employs a graph autoencoder to encode latent representations of the network's geometry, velocity fields, and concentration profiles. These latent space representations are then assembled into a global graph, which is subsequently used to predict system dynamics in the latent space via a trained graph latent space system dynamic model, inspired by the Neural Ordinary Differential Equations (Neural ODEs) concept. The integration of an autoencoder allows for the use of smaller graph neural network models with reduced training data requirements. Furthermore, the Neural ODE component effectively mitigates the issue of error accumulation commonly encountered in recurrent neural networks. The effectiveness of the GALDS model is demonstrated through results on eight unseen geometries and four abnormal transport examples, where our approach achieves mean relative error of 3% with maximum relative error <8% and demonstrates a 10-fold speed improvement compared to previous surrogate model approaches.

Figures

Figures reproduced from arXiv: 2507.10871 by the authors.

Figure 1
Figure 1. (A) A schematic overview of a typical surrogate modeling workflow using neurite transport as an example. (B) Time comparison of the [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. (A) Training process of all GALDS modules. The velocity autoencoder utilizes separated models for pipe and bifurcation geometries, [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Architecture of GALDS modules. (A) The autoencoder module takes the adjacency matrix [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Pipeline for the inference process of the GALDS model. This diagram outlines the four sequential steps, with each step clearly dis [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Predicted concentration fields within the neurite network geometries for NMO [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Nodal relative error maps for concentration predictions within the neurite network geometries for NMO [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Predicted velocity fields and corresponding relative error maps for two neurite geometries: NMO [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Spatially asymmetric attachment rates and resulting flow patterns in swelling and shrinking pipes. (A) Five training and one evaluation [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: GALDS prediction accuracy for spatially asymmetric attachment rate examples. (A-C) Results for the NMO [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Latent Dynamics Graph Convolutional Networks for model order reduction of parameterized time-dependent PDEs

    cs.LG 2026-01 conditional novelty 5.0 of 10

    LD-GCN couples an encoder-free latent-space neural ODE with a graph convolutional decoder, achieving accurate reduced-order modeling of time-dependent parameterized PDEs and detecting bifurcations from the latent traj...

Reference graph

Works this paper leans on

63 extracted references · 58 canonical work pages · cited by 1 Pith paper

  1. [48]

    A. Li, A. Barati Farimani, Y . J. Zhang, Deep learning of material transport in complex neurite networks, Scientific Reports 11 (1) (2021) 11280

    work page 2021

  2. [1]

    K. J. De V os, A. J. Grierson, S. Ackerley, C. C. Miller, Role of axonal transport in neurodegenerative diseases, Annu. Rev. Neurosci. 31 (1) (2008) 151–173

    work page 2008

  3. [2]

    Gunawardena, L

    S. Gunawardena, L. S. Goldstein, Polyglutamine diseases and transport problems: Deadly tra ffic jams on neuronal highways, Archives of Neurology 62 (1) (2005) 46–51. 22 Algorithm 1 Training Procedure for GALDS Require: Training dataset D ={(xi, ui, n0,i, n+,i)}N i=1 Ensure: Trained parametersθ(u) ϕ ,θ (u) ψ ,θ (n) ϕ ,θ (n) ψ Phase 1: Train Velocity Autoen...

    work page 2005

  4. [3]

    N. L. Kononenko, G. A. Claßen, M. Kuijpers, D. Puchkov, T. Maritzen, A. Tempes, A. R. Malik, A. Skalecka, S. Bera, J. Jaworski, et al., Ret- rograde transport of TrkB-containing autophagosomes via the adaptor AP-2 mediates neuronal complexity and prevents neurodegeneration, Nature Communications 8 (1) (2017) 14819

    work page 2017

  5. [4]

    Zhang, C

    H. Zhang, C. Zhang, J. Vincent, D. Zala, C. Benstaali, M. Sainlos, D. Grillo-Bosch, S. Daburon, F. Coussen, Y . Cho, et al., Modulation of AMPA receptor surface diffusion restores hippocampal plasticity and memory in Huntington’s disease models, Nature Communications 9 (1) (2018) 4272

    work page 2018

  6. [5]

    D. A. Smith, R. M. Simmons, Models of motor-assisted transport of intracellular particles, Biophysical Journal 80 (1) (2001) 45–68

    work page 2001

  7. [6]

    Friedman, G

    A. Friedman, G. Craciun, A model of intracellular transport of particles in an axon, Journal of Mathematical Biology 51 (2005) 217–246

    work page 2005

  8. [7]

    J. Wan, B. Steele, S. A. Spicer, S. Strohband, G. R. Feij ´oo, T. J. Hughes, C. A. Taylor, A one-dimensional finite element method for simulation-based medical planning for cardiovascular disease, Computer Methods in Biomechanics & Biomedical Engineering 5 (3) (2002) 195–206

    work page 2002

Show all 63 references
  1. [8]

    Y . J. Zhang, Geometric Modeling and Mesh Generation from Scanned Images, Chapman and Hall/CRC, 2016

    2016

  2. [9]

    Y . J. Zhang, Challenges and advances in image-based geometric modeling and mesh generation, in: Image-Based Geometric Modeling and Mesh Generation, Springer, 2013

    2013

  3. [10]

    L. M. Loew, J. C. Scha ff, The virtual cell: A software environment for computational cell biology, TRENDS in Biotechnology 19 (10) (2001) 401–406

    2001

  4. [11]

    T. J. Hughes, J. A. Cottrell, Y . Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh refinement, Computer Methods in Applied Mechanics and Engineering 194 (39-41) (2005) 4135–4195

    2005

  5. [12]

    Zhang, Y

    Y . Zhang, Y . Bazilevs, S. Goswami, C. L. Bajaj, T. J. Hughes, Patient-specific vascular NURBS modeling for isogeometric analysis of blood flow, Computer Methods in Applied Mechanics and Engineering 196 (29-30) (2007) 2943–2959

    2007

  6. [13]

    Zhang, X

    Y . Zhang, X. Liang, J. Ma, Y . Jing, M. J. Gonzales, C. Villongco, A. Krishnamurthy, L. R. Frank, V . Nigam, P. Stark, et al., An atlas- based geometry pipeline for cardiac Hermite model construction and di ffusion tensor reorientation, Medical Image Analysis 16 (6) (2012) 1130–1141

    2012

  7. [14]

    Urick, T

    B. Urick, T. M. Sanders, S. S. Hossain, Y . J. Zhang, T. J. Hughes, Review of patient-specific vascular modeling: Template-based isogeometric framework and the case for CAD, Archives of Computational Methods in Engineering 26 (2019) 381–404

    2019

  8. [15]

    Y . Yu, Y . J. Zhang, K. Takizawa, T. E. Tezduyar, T. Sasaki, Anatomically realistic lumen motion representation in patient-specific space–time isogeometric flow analysis of coronary arteries with time-dependent medical-image data, Computational Mechanics 65 (2020) 395–404

    2020

  9. [16]

    K. Qian, A. Pawar, A. Liao, C. Anitescu, V . Webster-Wood, A. W. Feinberg, T. Rabczuk, Y . J. Zhang, Modeling neuron growth using isogeometric collocation based phase field method, Scientific Reports 12 (1) (2022) 8120

    2022

  10. [17]

    K. Qian, G. O. Suarez, T. Nambara, T. Kanekiyo, A. S. Liao, V . A. Webster-Wood, Y . J. Zhang, Neurodevelopmental disorders modeling using isogeometric analysis, dynamic domain expansion and local refinement, Computer Methods in Applied Mechanics and Engineering 433 (2025) 117534

    2025

  11. [18]

    K. Qian, Y . J. Zhang, 3D neuron growth and neurodevelopmental disorder modeling based on truncated hierarchical B-splines with multi-level local refinements, Computer Methods in Applied Mechanics and Engineering 445 (2025) 118176

    2025

  12. [19]

    A. Li, X. Chai, G. Yang, Y . J. Zhang, An isogeometric analysis computational platform for material transport simulations in complex neurite networks, Molecular & Cellular Biomechanics 16 (2) (2019) 123–140

    2019

  13. [20]

    A. Li, Y . J. Zhang, Modeling material transport regulation and traffic jam in neurons using PDE-constrained optimization, Scientific Reports 12 (1) (2022) 3902

    2022

  14. [21]

    A. Li, Y . J. Zhang, Modeling intracellular transport and traffic jam in 3D neurons using PDE-constrained optimization, Journal of Mechanics 38 (2022) 44–59

    2022

  15. [22]

    D. Lu, D. Ricciuto, E fficient surrogate modeling methods for large-scale earth system models based on machine-learning techniques, Geosci- entific Model Development 12 (5) (2019) 1791–1807

    2019

  16. [23]

    Azarhoosh, M

    Z. Azarhoosh, M. I. Ghazaan, A review of recent advances in surrogate models for uncertainty quantification of high-dimensional engineering applications, Computer Methods in Applied Mechanics and Engineering 433 (2025) 117508

    2025

  17. [24]

    Raissi, P

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Machine learning of linear di fferential equations using gaussian processes, Journal of Computa- tional Physics 348 (2017) 683–693

    2017

  18. [25]

    R. K. Tripathy, I. Bilionis, Deep UQ: Learning deep neural network surrogate models for high dimensional uncertainty quantification, Journal of Computational Physics 375 (2018) 565–588

    2018

  19. [26]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional networks for biomedical image segmentation, in: Medical Image Computing and Computer-Assisted Intervention–MICCAI Part III 18, Springer, 2015, pp. 234–241

    2015

  20. [27]

    Thuerey, K

    N. Thuerey, K. Weißenow, L. Prantl, X. Hu, Deep learning methods for Reynolds-averaged Navier–Stokes simulations of airfoil flows, AIAA Journal 58 (1) (2020) 25–36

    2020

  21. [28]

    X. Guo, W. Li, F. Iorio, Convolutional neural networks for steady flow approximation, in: Proceedings of the 22nd ACM SIGKDD Interna- tional Conference on Knowledge Discovery and Data Mining, 2016, pp. 481–490

    2016

  22. [29]

    K. Qian, A. S. Liao, S. Gu, V . A. Webster-Wood, Y . J. Zhang, Biomimetic IGA neuron growth modeling with neurite morphometric features and CNN-based prediction, Computer Methods in Applied Mechanics and Engineering 417 (2023) 116213

    2023

  23. [30]

    Pfa ff, M

    T. Pfa ff, M. Fortunato, A. Sanchez-Gonzalez, P. Battaglia, Learning mesh-based simulation with graph networks, in: International Conference on Learning Representations, 2020

    2020

  24. [31]

    Sanchez-Gonzalez, J

    A. Sanchez-Gonzalez, J. Godwin, T. Pfa ff, R. Ying, J. Leskovec, P. Battaglia, Learning to simulate complex physics with graph networks, in: International Conference on Machine Learning, Proceedings of Machine Learning Research, 2020, pp. 8459–8468

    2020

  25. [32]

    Raissi, P

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics 378 (2019) 686–707

    2019

  26. [33]

    Hsieh, T.-H

    T.-Y . Hsieh, T.-H. Huang, A multiscale stabilized physics informed neural networks with weakly imposed boundary conditions transfer learning method for modeling advection dominated flow, Engineering with Computers 40 (5) (2024) 3353–3387

    2024

  27. [34]

    S. Wang, Y . Teng, P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM Journal 24 on Scientific Computing 43 (5) (2021) A3055–A3081

    2021

  28. [35]

    Krishnapriyan, A

    A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, M. W. Mahoney, Characterizing possible failure modes in physics-informed neural networks, Advances in Neural Information Processing Systems 34 (2021) 26548–26560

    2021

  29. [36]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Advances in Neural Information Processing Systems 30 (2017)

    2017

  30. [37]

    K. Qian, G. O. Suarez, T. Nambara, T. Kanekiyo, Y . J. Zhang, High-throughput machine learning framework for predicting neurite deteriora- tion using MetaFormer attention, Computer Methods in Applied Mechanics and Engineering 442 (2025) 118003

    2025

  31. [38]

    Feng, X.-P

    B. Feng, X.-P. Zhou, The novel graph transformer-based surrogate model for learning physical systems, Computer Methods in Applied Mechanics and Engineering 432 (2024) 117410

    2024

  32. [39]

    Zhongbo, P

    Y . Zhongbo, P. L. Hien, Pre-trained transformer model as a surrogate in multiscale computational homogenization framework for elastoplastic composite materials subjected to generic loading paths, Computer Methods in Applied Mechanics and Engineering 421 (2024) 116745

    2024

  33. [40]

    Hemmasian, A

    A. Hemmasian, A. B. Farimani, Multi-scale time-stepping of partial di fferential equations with transformers, Computer Methods in Applied Mechanics and Engineering 426 (2024) 116983

    2024

  34. [41]

    L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators, Nature Machine Intelligence 3 (3) (2021) 218–229

    2021

  35. [42]

    W. Li, M. Z. Bazant, J. Zhu, Phase-field DeepONet: Physics-informed deep operator neural network for fast simulations of pattern formation governed by gradient flows of free-energy functionals, Computer Methods in Applied Mechanics and Engineering 416 (2023) 116299

    2023

  36. [43]

    M. Zhu, S. Feng, Y . Lin, L. Lu, Fourier-DeepONet: Fourier-enhanced deep operator networks for full waveform inversion with improved accuracy, generalizability, and robustness, Computer Methods in Applied Mechanics and Engineering 416 (2023) 116300

    2023

  37. [44]

    J. He, S. Koric, S. Kushwaha, J. Park, D. Abueidda, I. Jasiuk, Novel DeepONet architecture to predict stresses in elastoplastic structures with variable complex geometries and loads, Computer Methods in Applied Mechanics and Engineering 415 (2023) 116277

    2023

  38. [45]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Fourier neural operator for parametric partial differential equations, arXiv preprint arXiv:2010.08895 (2020)

    2020 arXiv

  39. [46]

    Lehmann, F

    F. Lehmann, F. Gatti, M. Bertin, D. Clouteau, 3D elastic wave propagation with a factorized fourier neural operator (F-FNO), Computer Methods in Applied Mechanics and Engineering 420 (2024) 116718

    2024

  40. [47]

    X. Ye, H. Li, J. Huang, G. Qin, On the locality of local neural operator in learning fluid dynamics, Computer Methods in Applied Mechanics and Engineering 427 (2024) 117035

    2024

  41. [49]

    A. Li, Y . J. Zhang, Isogeometric analysis-based physics-informed graph neural network for studying traffic jam in neurons, Computer Methods in Applied Mechanics and Engineering 403 (2023) 115757

    2023

  42. [50]

    X. Wei, Y . J. Zhang, T. J. Hughes, Truncated hierarchical tricubicC0 spline construction on unstructured hexahedral meshes for isogeometric analysis applications, Computers & Mathematics with Applications 74 (9) (2017) 2203–2220

    2017

  43. [51]

    X. Wei, Y . Zhang, L. Liu, T. J. Hughes, Truncated T-splines: Fundamentals and methods, Computer Methods in Applied Mechanics and Engineering 316 (2017) 349–372

    2017

  44. [52]

    X. Wei, Y . J. Zhang, D. Toshniwal, H. Speleers, X. Li, C. Manni, J. A. Evans, T. J. Hughes, Blended B-spline construction on unstructured quadrilateral and hexahedral meshes with optimal convergence rates in isogeometric analysis, Computer Methods in Applied Mechanics and Eng...

    2018

  45. [53]

    de Boor, Subroutine package for calculating with B-Splines, in: Los Alamos Sci

    C. de Boor, Subroutine package for calculating with B-Splines, in: Los Alamos Sci. Lab., Los Alamos, NM, USA, 1971

    1971

  46. [54]

    Casquero, X

    H. Casquero, X. Wei, D. Toshniwal, A. Li, T. J. Hughes, J. Kiendl, Y . J. Zhang, Seamless integration of design and Kirchho ff–Love shell analysis using analysis-suitable unstructured T-splines, Computer Methods in Applied Mechanics and Engineering 360 (2020) 112765

    2020

  47. [55]

    X. Wei, X. Li, K. Qian, T. J. Hughes, Y . J. Zhang, H. Casquero, Analysis-suitable unstructured T-splines: Multiple extraordinary points per face, Computer Methods in Applied Mechanics and Engineering 391 (2022) 114494

    2022

  48. [56]

    Carson, R

    J. Carson, R. Van Loon, An implicit solver for 1D arterial network models, International Journal for Numerical Methods in Biomedical Engineering 33 (7) (2017) e2837

    2017

  49. [57]

    M. Poli, S. Massaroli, J. Park, A. Yamashita, H. Asama, J. Park, Graph neural ordinary differential equations, arXiv preprint arXiv:1911.07532 (2019)

    2019 arXiv

  50. [58]

    R. T. Chen, Y . Rubanova, J. Bettencourt, D. K. Duvenaud, Neural ordinary differential equations, Advances in Neural Information Processing Systems 31 (2018)

    2018

  51. [59]

    M. W. Gramlich, L. Conway, W. H. Liang, J. A. Labastide, S. J. King, J. Xu, J. L. Ross, Single molecule investigation of kinesin-1 motility using engineered microtubule defects, Scientific Reports 7 (1) (2017) 1–12

    2017

  52. [60]

    Bugiel, A

    M. Bugiel, A. Mitra, S. Girardo, S. Diez, E. Sch ¨affer, Measuring microtubule supertwist and defects by three-dimensional-force-clamp tracking of single kinesin-1 motors, Nano Letters 18 (2) (2018) 1290–1295

    2018

  53. [61]

    Fern ´andez, P

    P. Fern ´andez, P. A. Pullarkat, The role of the cytoskeleton in volume regulation and beading transitions in PC12 neurites, Biophysical Journal 99 (11) (2010) 3571–3579

    2010

  54. [62]

    Datar, J

    A. Datar, J. Ameeramja, A. Bhat, R. Srivastava, A. Mishra, R. Bernal, J. Prost, A. Callan-Jones, P. A. Pullarkat, The roles of microtubules and membrane tension in axonal beading, retraction, and atrophy, Biophysical Journal 117 (5) (2019) 880–891

    2019

  55. [63]

    M. S ¸. Aydın, S. Bay, E. N. Yi ˘git, C. ¨Ozg¨ul, E. K. O ˘guz, E. Y . Konuk, N. Ays ¸it, N. Cengiz, E. Erdo˘gan, A. Him, et al., Active shrinkage protects neurons following axonal transection, iScience 26 (10) (2023). 25

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.