REVIEW 4 major objections 5 minor 17 references
Convolutional Long Short-Term Memory Neural Networks Based Numerical Simulation of Flow Field
T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read An improved ConvLSTM predicts cylinder-wake flow fields with 41% lower mean-squared error and 37% fewer parameters than the standard model.
desk verdict Standard ConvLSTM tweak whose reported gains likely come from a random temporal split; fixable but not yet convincing. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central mechanism is the SE-Res3D-ConvLSTM module: a Squeeze-and-Excitation channel-attention block combined with a 3D residual network, placed ahead of a standard ConvLSTM, together with a Conv3D layer that adjusts feature dimensionality. The residual connections let the network train a deeper spatial feature extractor without vanishing-gradient problems; the channel attention re-weights feature maps channel by channel; and the ConvLSTM core handles the temporal recurrence. The paper credits this module for the simultaneous improvement in accuracy and reduction in parameter count and training time.
What would settle it
Train and evaluate the same two models on the same CFD data but split the timeline chronologically—train on the first 70% of timesteps, validate on the next 10%, and test on the final 20%—or test on a simulation at a different Reynolds number; if the improved model's MAE and MSE advantages shrink or reverse, the random-split experiment did not establish genuine predictive generalization.
Extended reading notes
Core claim
The paper's central claim is that an SE-residual-3D-ConvLSTM architecture extracts more temporal and spatial features from cylinder-wake flow fields than the standard ConvLSTM while requiring less computation. The author reports total parameter counts dropping from 495,457 to 310,385, training time dropping from 560 to 381 minutes, MAE dropping from 142 to 139, MSE dropping from 140 to 82, and SSIM rising from 0.71 to 0.84. These numbers are the evidence for the conclusion that attention mechanisms, residual networks, and 3D convolutions are beneficial additions to ConvLSTM for flow-field prediction. The author also notes that prediction accuracy decreases as the forecast horizon grows, a limitation of the present approach.
Load-bearing premise
The load-bearing premise is that randomly splitting one continuous, repeating simulation into training and test snapshots gives independent test samples, so the reported test error measures how well the model predicts unseen flow states rather than how well it memorizes a periodic sequence.
Editorial extensions
If this is right
- On the paper's evidence, the improved ConvLSTM predicts the next flow-field frame of the cylinder wake with a 41% lower MSE and an SSIM of 0.84 rather than 0.71.
- The improved model trains in roughly two-thirds of the time of the standard ConvLSTM on the same hardware, making repeated retraining cheaper.
- It uses about 185,000 fewer parameters, so it is lighter to embed in design-optimization loops that call the surrogate many times.
- The success of residual and attention modules on this 2D spatiotemporal task suggests these components transfer beyond image classification to flow-field forecasting.
Reading between the lines
- Inference beyond the paper: a chronological train/test split, rather than a random split of one periodic simulation, would directly test whether the reported gains reflect learned flow dynamics rather than memorization of a repeating sequence.
- Inference beyond the paper: because the CFD dataset covers a single Reynolds number and cylinder configuration, testing the same architecture at different Reynolds numbers or cylinder spacings would show whether the SE-residual-ConvLSTM block is a general spatiotemporal feature extractor.
- Inference beyond the paper: the same block could be carried to other image-like field-forecasting problems, such as weather-radar nowcasting or unsteady aerodynamics, since the input format—a sequence of 2D fields—is identical.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes an improved ConvLSTM network that combines residual modules, channel attention (SENet), and 3D convolutions for predicting the flow field in the wake of a circular cylinder. The authors generate a dataset from CFD simulations at Re=200, then compare the proposed model against a standard ConvLSTM in terms of parameter count, training time, MAE, MSE, and SSIM. The central claim, stated in the abstract and Section 4, is that the improved model extracts richer spatiotemporal features while having fewer parameters and shorter training time, with Table 1 reporting a 37% reduction in parameters, 37% reduction in training time, 41% reduction in MSE, and an SSIM increase from 0.71 to 0.84.
Significance. If the comparison were valid, the contribution would be a modest architectural improvement over a standard baseline in an application area where deep-learning surrogates for CFD are of interest. The paper provides a documented CFD dataset-generation procedure and a clear description of the proposed architecture. However, the evaluation is currently not trustworthy because of the random temporal split of a periodic time series, and the paper does not compare with any recent spatiotemporal prediction methods. The strength of the claim is therefore limited to a preliminary indication that the proposed modules can reduce parameter count; the claimed accuracy gains are not established by the presented evidence.
major comments (4)
- [§2.3 and §4, Table 1] The dataset is a single continuous time series of a periodic vortex-shedding flow sampled every 0.02 s. Partitioning it randomly with a random seed places temporally adjacent snapshots in both the training and test sets. Because adjacent snapshots are highly correlated and the flow is periodic, the test scores in Table 1 measure the model's ability to interpolate between or memorize training frames rather than to predict unseen future flow states. The statement in Section 5 that prediction accuracy decreases as the prediction time step increases reinforces this concern. The evaluation should be redone with a strict chronological split, for example training on the first 70% of the trajectory and validating/testing on later non-overlapping time intervals, and the authors should report whether the Table 1 improvements persist under that split.
- [§4, Table 1] All metrics are reported as single numbers from one run, without error bars, confidence intervals, or statistical tests. Because training time and accuracy of stochastic deep networks vary across runs, the reported 41% MSE reduction and 18% SSIM increase may not be significant. The authors should report the mean and standard deviation over at least five random seeds and state whether the differences are statistically significant, for example with a paired test across matched runs.
- [§4, Figs. 6 and 7] The prediction experiment is not specified precisely enough to interpret the metrics. The manuscript does not state how many past frames are used as input, how many future frames are predicted, or whether the prediction is single-step or recursive. The phrases 'frame 11' and 'frame 14' in Section 4 are undefined. The authors should specify the input/output sequence mapping, the number of timesteps, and the prediction horizon; otherwise the reported MAE, MSE, and SSIM values are not reproducible.
- [§2.2 and §4, Table 1] The numerical setup is described as two tandem circular cylinders, but the abstract and several passages refer to a single cylinder; this inconsistency makes the dataset description unclear. In addition, the agreement with published literature is only asserted qualitatively, without quantitative values such as Strouhal number or mean drag/lift coefficients. Finally, the comparison in Table 1 is only against the standard ConvLSTM; without other baseline models, the claim that the proposed model 'extracts more temporal and spatial features' is not contextualized against the broader state of the art.
minor comments (5)
- [Abstract and §1] There are typos and inconsistent terminology: 'flow filed' should be 'flow field'; 'Con-vLSTM' appears in the abstract while 'ConvLSTM' is used elsewhere; and 'ReNet' in Section 3.4 should be 'ResNet' to match the residual-network terminology.
- [§4, figures] The figure numbering is inconsistent: Section 4 refers to 'Fig. 5' for prediction results, but Fig. 5 in Section 3.4 is the network structure. The prediction figures should be renumbered consistently.
- [§4, Eqs. (7)-(9)] The evaluation metrics are dimensionally ambiguous because the flow-field variables and any normalization are not specified. Please state what quantity is being compared (e.g., velocity magnitude), the data range, and how the errors are computed in physical or normalized units.
- [§5] The statement that dropout produced only trivial improvement is anecdotal; if mentioned, please provide the numerical results for that experiment.
- [§2.3] The dataset size is not reported. Please specify the number of snapshots, the sequence length used for training, and the number of samples excluded from the initial transient phase.
Circularity Check
The predictive evaluation is undermined by random temporal splitting, so the reported accuracy gains are partly forced by construction; parameter and training-time reductions are not circular.
-
fitted input called prediction
[Section 2.3 (Dataset) and Section 4 (Table 1) and Section 5 (Conclusions)]
"Although the numerical simulation used a time step of 0.001s, adopting this interval for dataset construction would result in negligible observable changes in the wake flow field behind the cylinder. ... the sampling interval is set as 0.02s. ... The dataset was partitioned using a random seed generation method, with the following allocation: 70% of the data was assigned to the training set; 10% to the validation set; 20% to the test set."
The data come from one continuous periodic vortex-shedding simulation, so snapshots 0.02 s apart are near-duplicates. A random 70/10/20 split interleaves temporally adjacent frames across train and test rather than holding out future time. Network weights are fitted to frames surrounding the test frames, so Table 1's MAE/MSE/SSIM measure interpolation/memorization inside the training trajectory, not forecasting unseen states. The abstract and Section 4 nonetheless present these numbers as evidence that the improved ConvLSTM can extract more temporal and spatial features; that predictive claim is forced by the split construction, not demonstrated.
full rationale
No self-citation chain, imported uniqueness theorem, or definitional equivalence appears in the paper; the architecture comparison and the parameter/time reductions in Table 1 are genuine properties of the two models and are not circular. The circularity is confined to the evaluation of predictive accuracy: random splitting of a single continuous periodic simulation makes the test set statistically dependent on the training set, so the reported 41% MSE reduction and SSIM increase do not establish predictive generalization. The paper's own conclusion concedes that accuracy degrades as the prediction time step increases, which is consistent with the absence of a temporal holdout. I score 6 rather than 7 or 8 because the central efficiency claims (fewer parameters, shorter training time) remain independently supported by the architecture, while the accuracy-based 'prediction' claim is partially reduced to in-trajectory interpolation.
Assumptions & free parameters
free parameters (2)
- Model hyperparameters (kernel sizes, layer counts, learning rate, batch size, optimizer, loss weighting) =
not reported
- Random seed for train/validation/test split =
not reported
assumptions (3)
- domain assumption The CFD solution (ANSYS, dynamic mesh, UDF) provides ground-truth flow fields
- ad hoc to paper Randomly splitting a continuous time series into train/validation/test yields independent samples
- domain assumption Sampling every 0.02s from the 0.001s CFD time step captures the flow dynamics without aliasing
Cite this review
Pith. "Pith review of Convolutional Long Short-Term Memory Neural Networks Based Numerical Simulation of Flow Field." pith.science (2026). https://pith.science/paper/JM2MU7VA
@misc{pith2026250515533,
author = {Pith},
title = {Pith review of: Convolutional Long Short-Term Memory Neural Networks Based Numerical Simulation of Flow Field},
year = {2026},
howpublished = {\url{https://pith.science/paper/JM2MU7VA}},
note = {Machine review of arXiv:2505.15533}
}
read the original abstract
Computational Fluid Dynamics (CFD) is the main approach to analyzing flow field. However, the convergence and accuracy depend largely on mathematical models of flow, numerical methods, and time consumption. Deep learning-based analysis of flow filed provides an alternative. For the task of flow field prediction, an improved Convolutional Long Short-Term Memory (Con-vLSTM) Neural Network is proposed as the baseline network in consideration of the temporal and spatial characteristics of flow field. Combining dynamic mesh technology and User-Defined Function (UDF), numerical simulations of flow around a circular cylinder were conducted. Flow field snapshots were used to sample data from the cylinder's wake region at different time instants, constructing a flow field dataset with sufficient volume and rich flow state var-iations. Residual networks and attention mechanisms are combined with the standard ConvLSTM model. Compared with the standard ConvLSTM model, the results demonstrate that the improved ConvLSTM model can extract more temporal and spatial features while having fewer parameters and shorter train-ing time.
Reference graph
Works this paper leans on
-
[1]
Lee S, You D.: Prediction of laminar vortex shedding over a cylinder using deep learning. arXiv preprint arXiv: 1712. 07854 (2017)
work page 2017
-
[2]
Lee S, You D.: Data-driven prediction of unsteady flow over a circular cylinder using deep learning. Journal of Fluid Mechanics. 879, 217–254 (2019)
work page 2019
-
[3]
Kastner P, Dogan T.: A GAN -based surrogate model for instantaneous urban wind flow prediction. Building and Environment. 242, 242 (2023)
work page 2023
-
[4]
Doctor Thesis, The University of North Carolina at Charlotte, (2021)
Rodriguez F J C.: Generative Machine Learning Models for Airflow Prediction of Archi- tectural Spaces. Doctor Thesis, The University of North Carolina at Charlotte, (2021)
work page 2021
-
[5]
Hu C, Kikumoto H, Zhang B.: Fast estimation of airflow distribution in an urban model using generative adversarial networks with limited sensing data. Building and Environ- ment. 249, 111120 (2024)
work page 2024
-
[6]
Gori, M. Monfardini, G. Scarselli, F.: A new model for learning in graph domains. In: Pro- ceedings of 2005 IEEE International Joint Conference on Neural Networks, pp. 729-734. Montreal, QC, Canada (2005)
work page 2005
-
[7]
Journal of Machine Learning Research
Chami I, Abu -El-Haija S, Perozzi B.: Machine learning on graphs: A model and compre- hensive taxonomy. Journal of Machine Learning Research. 2389, 1–64 (2022)
work page 2022
-
[8]
arXiv preprint arXiv:1609.02907
Kipf T N, Welling M.: Semi -supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907. (2016)
arXiv 2016
Show all 17 references
-
[9]
Computational Mechanics
Bhatnagar S, Afshar Y, Pan S.: Prediction of aerodynamic flow fields using convolutional neural networks. Computational Mechanics. 64, 525–545 (2019)
2019
-
[10]
arXiv preprint arXiv:2004.08826
Ribeiro M D, Rehman A, Ahmed S.: DeepCFD: Efficient steady -state laminar flow ap- proximation with deep convolutional neural networks. arXiv preprint arXiv:2004.08826. (2020)
2020 arXiv
-
[11]
Mathematics
Abucide-Armas A, Portal -Porras K, Fernandez -Gamiz U.: A data augmentation -based technique for deep learning applied to CFD simulations. Mathematics. 916, 1843 (2021)
2021
-
[12]
Journal of Marine Science and Engineering
Abucide-Armas A, Portal -Porras K, Fernandez -Gamiz U.: Convolutional neural network predictions for unsteady Reynolds Averaged Navier Stokes Based numerical simulations. Journal of Marine Science and Engineering. 112, 239 (2023)
2023
-
[13]
Mathematics
Portal-Porras K, Fernandez-Gamiz U, Ugarte-Anero A.: YOLO-Fish: Alternative artificial neural network structures for turbulent flow velocity field prediction. Mathematics. 96, 1939 (2021)
2021
-
[14]
Computer graphics forum
Wiewel S, Becher M, Thuerey N.: Latent space physics: Towards learning the temporal evolution of fluid flow. Computer graphics forum. 38(2), 71–82 (2019)
2019
-
[15]
Applied Sciences
Ko J, Choi W, Lee S.: Internal flow prediction in arbitrary shaped channel using stream- wise bidirectional LSTM. Applied Sciences. 13(20), 11481 (2023)
2023
-
[16]
Proceed- ings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering
Wu Y, Dai Y, Yang C.: Unsteady and nonlinear aerodynamic prediction of airfoil under- going large-amplitude pitching oscillation based on gated recurrent unit network. Proceed- ings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering. 2372, 270...
2023
-
[17]
In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp
He K, Zhang X, Ren S.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770-778. Las Vegas, NV, USA (2016)
2016
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.