Pith. sign in

REVIEW 4 major objections 6 minor 38 references

FR-Mamba: Time-Series Physical Field Reconstruction Based on State Space Model

T0 review · 4 major / 6 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 A hybrid of the Mamba state-space model and Fourier neural operators reconstructs unsteady vorticity fields from 16 sparse sensors with roughly one-third the error of the previous best method on the cylinder-flow benchmark.

desk verdict Plausible Mamba+FNO architecture, but the L notation and train split contradictions make the reported benchmark results unreproducible as written. read the letter →

arxiv 2505.16083 v1 pith:TRYN4QC2 submitted 2025-05-21 cs.LG

Jiahuan Long , Wenzhe Zhang , Ning Wang , Tingsong Jiang , Wen Yao This is my paper
classification cs.LG
keywords PhysicalfieldreconstructionStatespacemodelMambaFourierneuraloperatorSparsesensormeasurementsVorticityFlowaroundacylinderLong-rangetemporaldependencies
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 sets out to establish that time-series physical field reconstruction—recovering full vorticity fields from a handful of sensor measurements, a task central to fluid dynamics and thermodynamics—is substantially more accurate when the model exploits the whole history of measurements rather than a single snapshot. It proposes FR-Mamba, a dual-branch architecture whose temporal branch uses the Mamba state-space model to encode long-range dependencies in linear time and whose spatial branch uses a Fourier neural operator to recover structure lost when sensor data are flattened into a sequence. On the $Re=100$ flow-around-a-cylinder benchmark with 16 sensors, the paper reports mean absolute error $0.31\times10^{-4}$ and maximum absolute error $0.10\times10^{-2}$, about one-third of the previous best results by PFR-Transformer. If those numbers are right, long-sequence temporal modeling is a decisive and cheap ingredient for reconstructing time-evolving physical fields.

What carries the argument

The load-bearing object is the FNO-Mamba block, a dual-branch module with a temporal branch and a spatial branch. The temporal branch is a selective state-space model (Mamba): embeddings produce per-step control matrices $B$ and $C$ and a learnable time step $\Delta$; the continuous transition matrix $A$ is discretized by zero-order hold; and the hidden state is updated recurrently but computed in parallel across feature dimensions, so long-range dependencies are encoded with linear time complexity. The spatial branch is a one-dimensional Fourier neural operator: each sensor channel is transformed with a discrete Fourier transform, only the first $m$ frequency modes are retained, each mode is multiplied by a learnable complex weight matrix $W_k$, and an inverse transform returns a global spatial representation. A gating mechanism fuses the two branch outputs, residual connections stack the blocks, and a 2D Fourier operator with a residual connection refines the final reconstructed field. The ablations show what this machinery is doing: without either Fourier operator, Max-AE is $0.355\times10^{-2}$; the 1D operator alone cuts it to $0.125\times10^{-2}$, and the full model reaches $0.108\times10^{-2}$.

What would settle it

Re-running the identical benchmark with 16 sensors on the $Re=100$ cylinder flow, a single fixed training split (for instance 400 training snapshots), and one consistent sequence length should reproduce FR-Mamba's average MAE near $0.31\times10^{-4}$ and Max-AE near $0.10\times10^{-2}$, with both below PFR-Transformer's $0.91\times10^{-4}$ and $0.53\times10^{-2}$ on every test interval; a reversal of that ordering, or a large shift in the absolute errors, would settle the claim against the paper.

Watch

Extended reading notes

Core claim

The central claim is that FR-Mamba is the first state-space-model-based physical field reconstruction method and that it outperforms all compared baselines—SVR-POD, LSTM-POD, Voronoi UNet, Shallow Decoder, Voronoi CNN, Gappy MLP, and PFR-Transformer—on both mean and maximum absolute error for time-series cylinder-flow reconstruction. The authors' key assertion is that temporal context, not just spatial encoding, is what earlier methods miss, and that a selective state-space model supplies that context at linear cost. The reported average errors are $0.31\times10^{-4}$ MAE and $0.10\times10^{-2}$ Max-AE across the five test intervals, compared with $0.91\times10^{-4}$ and $0.53\times10^{-2}$ for PFR-Transformer, with the error staying nearly flat from the first interval to the last. The ablation evidence is part of the claim: removing both Fourier operators raises MAE to $0.397\times10^{-4}$ and Max-AE to $0.355\times10^{-2}$; adding the 1D operator alone brings them to $0.315\times10^{-4}$ and $0.125\times10^{-2}$; adding the 2D operator reaches the final $0.310\times10^{-4}$ and $0.108\times10^{-2}$.

Load-bearing premise

The whole comparison rests on the experiment being run exactly as described—a fixed 16-sensor layout on the $Re=100$ cylinder-flow dataset, one unambiguous training/test split, and the same meaning of sequence length in the temporal branch and the reconstruction head; if any of those conventions differ, the reported error numbers cannot be reproduced or compared.

Editorial extensions

If this is right

  • A pure Mamba without spatial encoding is much weaker: removing both Fourier operators raises MAE from $0.310\times10^{-4}$ to $0.397\times10^{-4}$ and Max-AE from $0.108\times10^{-2}$ to $0.355\times10^{-2}$, so frequency-domain spatial modeling is a necessary complement to the state-space model.
  • The model's error stays nearly flat across the five test intervals (MAE 0.30–0.32, Max-AE 0.10–0.11), so it does not appear to accumulate error on long sequences the way recurrent baselines do.
  • Because the SSM branch runs in linear time, scaling to longer sensor histories should be cheaper than attention-based physical field reconstruction approaches; the paper frames this as the main practical advantage.
  • The best configuration found is moderate depth with a compact hidden dimension ($N_{layer}=10$, $D=8$); deeper and wider models degrade, which the paper attributes to overfitting under limited training data.
  • The reconstructed quantities are vorticity fields on a fixed 2D grid from a fixed 16-sensor layout, so the method is directly applicable to other physical-field reconstruction problems with the same geometric setup and time-series sensor inputs.

Reading between the lines

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

  • Beyond the paper: the same dual-branch design should transfer to other sparse-sensor PDE fields (pressure, temperature, concentration) whenever the sensor layout is fixed and the output lives on a regular grid, because neither branch assumes cylinder-specific physics.
  • Beyond the paper: the flat T1–T5 error suggests the model does not compound error over long sequences within the tested window, but the paper does not test beyond the 1000-snapshot test set; running the trained model on much longer rollouts would show where accumulation eventually begins.
  • Beyond the paper: the sharp degradation at $N_{layer}=20$ hints that limited training data, not architecture capacity, is the active constraint on this benchmark, so the reported gap over baselines might grow or shrink with more training snapshots.
  • Beyond the paper: pairing the deterministic MAE-trained decoder with a probabilistic head would let the model output uncertainty alongside the field, which is what sensor-based reconstruction needs for engineering decisions.
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. FR-Mamba is a dual-branch neural architecture for time-series physical field reconstruction from sparse sensor measurements. The temporal branch uses Mamba/SSM to capture long-range dependencies, the spatial branch uses a 1D Fourier Neural Operator, a gating mechanism fuses the two, and a 2D FNO refines the output. The authors evaluate on two-dimensional cylinder flow at Re=100 with 16 sensors and 192×112 snapshots, reporting state-of-the-art MAE and Max-AE against seven baselines over five time intervals, together with ablations on network depth, hidden dimension, and the FNO components.

Significance. The problem is timely and the proposed FNO+Mamba combination is a plausible direction for long-sequence physical field reconstruction. If the reported performance were reproducible, a three-fold reduction in MAE over the previous best method would be a useful empirical contribution. However, the paper currently provides no code, no statistical uncertainty estimates, and the experimental evidence rests on a single dataset; more importantly, the architecture as written cannot produce the time-series outputs that the reported T1–T5 metrics claim to measure. The significance is therefore conditional on resolving the internal inconsistencies described below.

major comments (4)
  1. [Section 3.4.2, Eq. (22) vs Section 3.3.1, Eqs. (8)-(9)] Section 3.4.2 defines L as the spatial grid size (L = H × W) when reshaping the FFN output from R^{B×L×C} to R^{B×H×W×C}, while Section 3.3.1 and Eqs. (8)-(9) use L as the number of temporal steps in the SSM. Because the same symbol denotes both quantities, the architecture as written either produces a single spatial field (if L = H × W) or cannot reshape its temporal output to a 2D field (if L is the temporal length). In neither case can it output the sequence of fields \hat{u}_1,...,\hat{u}_T required by Section 3.1, making the T1–T5 reconstruction metrics in Tables 1 and 2 unreproducible without additional specification of how successive field snapshots are generated.
  2. [Section 4.1.1 vs Figure 6] Section 4.1.1 states that "the first 400 snapshots are used for training, and the last 1000 snapshots are used for testing," but Figure 6 is captioned "4000 training and 1000 validation snapshots." The five evaluation intervals T1–T5 in Section 4.2.1 are defined as snapshots 3001–3400 through 4601–5000, which overlap the training portion under the Figure 6 description and are not aligned with the last-1000 test set under the text description. This prevents a reader from knowing which snapshots were actually used for training and testing and does not exclude the risk of test-set contamination.
  3. [Tables 1, 2, and 3] Tables 1 and 2 report a single MAE and Max-AE value per method and interval, with no error bars, number of seeds, or statistical tests, and no code is provided. Given that Table 3 shows the model's own MAE varying from 0.310 to 0.625 as Nlayer and D change, and from 0.310 to 0.397 across the FNO ablations in Table 4, the claimed superiority over PFR-Transformer cannot currently be distinguished from training noise, hyperparameter selection effects, or differences in experimental protocol.
  4. [Section 3.3.1, Eqs. (6)-(7)] Equations (6)-(7) are dimensionally inconsistent: Δ^{(i)} ∈ R^{B×L×M}, A is described as a "shared state transition matrix," and the hidden state is h(t_k) ∈ R^N. The product Δ A in \bar{A} = exp(Δ A) and the inverse A^{-1} in Eq. (7) are therefore undefined unless a specific shape for A is supplied. This prevents an independent implementation of the SSM branch and affects the central architecture claim.
minor comments (6)
  1. [Section 3.3.2, Eq. (16)] The normalization factor 1/d in the inverse 1D DFT appears to be a typo; an inverse DFT over the m retained frequency modes should use a normalization consistent with the forward transform in Eq. (13).
  2. [Section 3.3.1, Eq. (10)] The convolution kernel K is written with C and B rather than the discretized versions \bar{B} and C, and the expression is not directly derived from Eqs. (6)-(9); please correct the notation.
  3. [Reference [18]] Reference [18] contains a placeholder DOI (10.1109/TPAMI.2023.1234567); please verify the citation and provide the correct bibliographic details.
  4. [Table 3 caption] The caption contains a typo: "Varing" should be "Varying."
  5. [Section 3.4.2] The text says the output channels C correspond to "e.g., velocity components," but the task defined in Section 3.1 is vorticity reconstruction; please clarify the number and physical meaning of the output channels.
  6. [Overall manuscript] The paper contains no code or data availability statement; adding these would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's contribution is an empirical architecture comparison, not a derivation that reduces to its inputs.

full rationale

The paper's central claim is architectural and experimental: FR-Mamba, a hybrid of Mamba (SSM) and FNO, is trained on sensor-to-field supervised pairs and compared against published baselines on a standard cylinder-flow dataset. I find no derivation chain in which a prediction is defined as its own input or in which a fitted parameter is relabeled a prediction. The method sections (Sections 3.3-3.4) build the model from standard components (Mamba, FNO, gating, FFN), and the loss (Eq. 25) is ordinary supervised MAE; no equation or parameter is fit to the test intervals T1-T5 and then reported as a prediction. Citations to prior works, including earlier works by the same research group (Gappy MLP [30] and the dataset reference in Section 4.1.1), are used only as baselines or data sources, not as justifications of FR-Mamba's effectiveness, so they are not load-bearing self-citations. Two internal inconsistencies exist - the symbol L denotes temporal sequence length in Section 3.3.1 but spatial grid size H×W in Section 3.4.2 (Eq. 22), and the training/validation split is described as 400 snapshots in the text but 4000 in Figure 6 - but these are reproducibility/correctness defects, not circularity: neither makes the reported result true by construction. Per the hard rules, no circular step can be quoted with a specific reduction, so the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim depends on standard mathematical machinery (Fourier transforms, ZOH discretization, state-space recursions) and on the domain assumption that 16 vorticity sensors determine the cylinder-wake field. The reported hyperparameters (10 layers, 8 hidden units) are selected on the validation set; the number of retained Fourier modes is not reported. No new physical entities are introduced.

free parameters (5)
  • Number of FNO-Mamba layers Nlayer = 10
    Selected via ablation in Table 3 as the best depth; D=8 gives MAE 0.310e-4.
  • Hidden state dimension D = 8
    Selected via ablation in Table 3; D=8 with Nlayer=10 is the reported best configuration.
  • Number of Fourier layers T = 4
    Fixed by design in Section 3.3.2, chosen without reported sensitivity analysis.
  • Number of retained Fourier modes m = not reported
    Section 3.3.2 says only the first m components are retained, but m is never specified, so it is an unreported free choice.
  • Learnable time-step parameter Delta and bias gamma = trained
    Used in SSM discretization (Eq. 5-7); fit to training data as part of the model.
assumptions (3)
  • domain assumption The incompressible Navier-Stokes simulation at Re=100 provides ground-truth vorticity fields, and 16 uniform sensors are sufficient to reconstruct the full field.
    Section 4.1.1 assumes the 16 sensor readings carry enough information for global reconstruction; the paper does not justify sensor count or placement.
  • standard math The Fourier transform and zero-order hold discretization are valid and numerically stable for these data.
    Sections 3.3.1 and 3.3.2 rely on ZOH and DFT without numerical stability analysis.
  • domain assumption Mean absolute error is the appropriate objective and metric for reconstruction quality.
    Section 4.1.2 defines MAE as the loss, but no discussion of perceptual or physical fidelity metrics is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FR-Mamba: Time-Series Physical Field Reconstruction Based on State Space Model." pith.science (2026). https://pith.science/paper/TRYN4QC2

@misc{pith2026250516083,
  author       = {Pith},
  title        = {Pith review of: FR-Mamba: Time-Series Physical Field Reconstruction Based on State Space Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TRYN4QC2}},
  note         = {Machine review of arXiv:2505.16083}
}
read the original abstract

Physical field reconstruction (PFR) aims to predict the state distribution of physical quantities (e.g., velocity, pressure, and temperature) based on limited sensor measurements. It plays a critical role in domains such as fluid dynamics and thermodynamics. However, existing deep learning methods often fail to capture long-range temporal dependencies, resulting in suboptimal performance on time-evolving physical systems. To address this, we propose FR-Mamba, a novel spatiotemporal flow field reconstruction framework based on state space modeling. Specifically, we design a hybrid neural network architecture that combines Fourier Neural Operator (FNO) and State Space Model (SSM) to capture both global spatial features and long-range temporal dependencies. We adopt Mamba, a recently proposed efficient SSM architecture, to model long-range temporal dependencies with linear time complexity. In parallel, the FNO is employed to capture non-local spatial features by leveraging frequency-domain transformations. The spatiotemporal representations extracted by these two components are then fused to reconstruct the full-field distribution of the physical system. Extensive experiments demonstrate that our approach significantly outperforms existing PFR methods in flow field reconstruction tasks, achieving high-accuracy performance on long sequences.

Figures

Figures reproduced from arXiv: 2505.16083 by the authors.

Figure 1
Figure 1. The framework of our FR-Mamba. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. The overview of temporal feature extraction branch. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. The overall calculation process of SSM. As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The overview of spatial feature extraction branch. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: The feature fusion of our FR-Mamba. E˜ (i) F = X (i) Emb2 ⊙ E (i) T + X (i) Emb2 ⊙ E (i) S , (20) E (i) F = E˜ (i) F + E (i−1) F , (21) where ⊙ denotes the Hadamard product. This design enables adaptive adjustment of spatiotemporal feature importance for better transie…
Figure 6
Figure 6. Figure 6: Structure of the time-resolved flow field dataset with 4000 training and 1000 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Visualization of different regions in the flow field. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: The qualitative comparison of our FR-Mamba with other representative meth [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Visualization of prediction results and corresponding error of our FR-Mamba. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 34 canonical work pages

  1. [1]

    J. Yu, J. S. Hesthaven, Flowfield reconstruction method using artificial neural network, Aiaa Journal 57 (2) (2019) 482–498

    work page 2019

  2. [2]

    D. Shu, Z. Li, A. B. Farimani, A physics-informed diffusion model for high-fidelity flow field reconstruction, Journal of Computational Physics 478 (2023) 111972

    2023

  3. [3]

    X. Deng, M. Guo, H. Chen, Y. Tian, J. Le, H. Zhang, Dual-path flow field reconstruction for a scramjet combustor based on deep learning, Physics of Fluids 34 (9) (2022)

    work page 2022

  4. [4]

    Z. Luo, L. Wang, J. Xu, Z. Wang, M. Chen, J. Yuan, A. C. Tan, Recon- struction of missing flow field from imperfect turbulent flows by machine learning, Physics of Fluids 35 (8) (2023)

    work page 2023

  5. [5]

    Q. D. Nguyen, W. Lu, L. Chan, A. Ooi, C. Lei, A state-of-the-art review of flows past confined circular cylinders, Physics of Fluids 35 (7) (2023). 25

    work page 2023

  6. [6]

    R. Das, A simulated annealing-based inverse computational fluid dy- namics model for unknown parameter estimation in fluid flow problem, International Journal of Computational Fluid Dynamics 26 (9-10) (2012) 499–513

    work page 2012

  7. [7]

    H. Zhou, Y. C. Soh, C. Jiang, X. Wu, Compressed representation learn- ing for fluid field reconstruction from sparse sensor observations, in: 2015 International Joint Conference on Neural Networks (IJCNN), IEEE, 2015, pp. 1–6

    work page 2015

  8. [8]

    Loiseau, B

    J.-C. Loiseau, B. R. Noack, S. L. Brunton, Sparse reduced-order mod- elling: sensor-based dynamics to full-state estimation, Journal of Fluid Mechanics 844 (2018) 459–490

    work page 2018

Show all 38 references
  1. [9]

    Z. Deng, Y. Chen, Y. Liu, K. C. Kim, Time-resolved turbulent veloc- ity field reconstruction using a long short-term memory (lstm)-based artificial intelligence framework, Physics of Fluids 31 (7) (2019)

    2019

  2. [10]

    Sherstinsky, Fundamentals of recurrent neural network and long short-term memory network, Physica D: Nonlinear Phenomena 404 (2020) 132306

    A. Sherstinsky, Fundamentals of recurrent neural network and long short-term memory network, Physica D: Nonlinear Phenomena 404 (2020) 132306

    2020

  3. [11]

    J. E. Santos, Z. R. Fox, A. Mohan, D. O’Malley, H. Viswanathan, N. Lubbers, Development of the senseiver for efficient field reconstruc- tion from sparse observations, Nature Machine Intelligence 5 (11) (2023) 1317–1325

    2023

  4. [12]

    A. Gu, K. Goel, C. R´ e, Efficiently modeling long sequences with struc- tured state spaces, in: Proceedings of the International Conference on Learning Representations (ICLR), 2022

    2022

  5. [13]

    M. Poli, N. Mishra, F. Becattini, J. Gall, A. Geiger, A hypothesis on the inductive bias of ssms, in: Proceedings of the International Conference on Learning Representations (ICLR), 2023

    2023

  6. [14]

    A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces, in: Proceedings of the First Conference on Language Mod- eling (COLM), 2024

    2024

  7. [15]

    Hatamizadeh, J

    A. Hatamizadeh, J. Kautz, Mambavision: A hybrid mamba-transformer vision backbone, arXiv preprint arXiv:2407.08083 (2024). 26

    2024 arXiv

  8. [16]

    X. Li, J. Wang, Y. Liu, Nlp-mamba: Enhancing long text understanding with state space models, Journal of Natural Language Engineering 29 (4) (2023) 567–580

    2023

  9. [17]

    Umatani, T

    R. Umatani, T. Imai, K. Kawamoto, S. Kunimasa, Time series clustering with an em algorithm for mixtures of linear gaussian state space models, Pattern Recognition 138 (2023) 109375

    2023

  10. [18]

    W. Gu, L. Zhang, M. Chen, Vision-mamba: Efficient state space models for real-time video processing, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (7) (2023) 1234–1245.doi:10.1109/TPAMI. 2023.1234567

    2023

  11. [19]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. M. Stuart, A. Anandkumar, Fourier neural operator for parametric partial differential equations, in: Proceedings of the International Conference on Learning Representations (ICLR), 2020

    2020

  12. [20]

    Pathak, S

    J. Pathak, S. Subramanian, P. Harrington, S. Raja, M. Brewster, D. Hall, M. Henaff, R. Keisler, D. Medina, R. Walters, et al., Fourcast- net: A global data-driven high-resolution weather model using adaptive fourier neural operators, in: Advances in Neural Information Processi...

    2022

  13. [21]

    Y. Wang, H. Wu, J. Zhang, Z. Gao, J. Wang, P. S. Yu, M. Long, Predrnn: A recurrent neural network for spatiotemporal predictive learning, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (2) (2022) 2208–2225

    2022

  14. [22]

    Huang, S

    Y. Huang, S. Zhu, S. Chen, Deep learning-driven super-resolution recon- struction of two-dimensional explosion pressure fields, Journal of Build- ing Engineering 78 (2023) 107620

    2023

  15. [23]

    G. Y. Tian, A. Sophian, D. Taylor, J. Rudlin, Multiple sensors on pulsed eddy-current detection for 3-d subsurface crack assessment, IEEE Sen- sors Journal 5 (1) (2005) 90–96

    2005

  16. [24]

    Fukami, Kai, Maulik, Romit, Ramachandra, Nesar, Fukagata, Koji, Taira, Kunihiko, Global field reconstruction from sparse sensors with Voronoi tessellation-assisted deep learning, Nature Machine Intelligence 3 (11) (2021) 945–951. 27

    2021

  17. [25]

    X. Peng, X. Li, Z. Gong, X. Zhao, W. Yao, A deep learning method based on partition modeling for reconstructing temperature field, International Journal of Thermal Sciences 182 (2022) 107802

    2022

  18. [26]

    X. Chen, Z. Gong, X. Zhao, W. Zhou, W. Yao, A machine learning surrogate modeling benchmark for temperature field reconstruction of heat source systems, Science China Information Sciences 66 (5) (2023) 152203

    2023

  19. [27]

    H. Su, J. Jiang, A. Wang, W. Zhuang, X.-H. Yan, Subsurface tempera- ture reconstruction for the global ocean from 1993 to 2020 using satellite observations and deep learning, Remote Sensing 14 (13) (2022) 3198

    2022

  20. [28]

    Gherlone, P

    M. Gherlone, P. Cerracchio, M. Mattone, M. Di Sciuva, A. Tessler, Shape sensing of 3d frame structures using an inverse finite element method, International Journal of Solids and Structures 49 (22) (2012) 3100–3112

    2012

  21. [29]

    Y. Gu, L. Wang, W. Chen, C. Zhang, X. He, Application of the meshless generalized finite difference method to inverse heat source problems, International Journal of Heat and Mass Transfer 108 (2017) 721–729

    2017

  22. [30]

    Zhao, X, Chen, X, Gong, Z, et al., A hybrid method based on proper or- thogonal decomposition and deep neural networks for flow and heat field reconstruction, Expert Systems with Applications 247 (2024) 123137

    2024

  23. [31]

    Y. Xia, W. Yao, X. Zheng, Z. Gong, Reliability analysis of heat source layout temperature field prediction considering uncertainty in deep neu- ral network surrogate models, Quality and Reliability Engineering In- ternational 39 (5) (2023) 1775–1795

    2023

  24. [32]

    N. B. Erichson, et al., Shallow decoder networks for fluid flow recon- struction, Journal of Fluid Mechanics 935 (2021) A10

    2021

  25. [33]

    A. Gu, I. Johnson, K. Goel, K. Saab, T. Dao, A. Rudra, C. R´ e, Com- bining recurrent, convolutional, and continuous-time models with linear state space layers, Advances in neural information processing systems 34 (2021) 572–585. 28

    2021

  26. [34]

    J. T. Smith, A. Warrington, S. W. Linderman, Simplified state space layers for sequence modeling, in: Proceedings of the Eleventh Interna- tional Conference on Learning Representations (ICLR), 2023

    2023

  27. [35]

    D. Y. Fu, T. Dao, K. K. Saab, A. W. Thomas, A. Rudra, C. R´ e, Hungry hungry hippos: Towards language modeling with state space models, arXiv preprint arXiv:2212.14052 (2022)

    2022 arXiv

  28. [36]

    K. C. Pohlmann, Principles of digital audio, McGraw-Hill Professional, 2000

    2000

  29. [37]

    X. Zhao, Z. Gong, X. Chen, W. Yao, Y. Zhang, A unified framework of deep neural networks and gappy proper orthogonal decomposition for global field reconstruction, in: 2023 International Joint Conference on Neural Networks (IJCNN), IEEE, 2023, pp. 1–8

    2023

  30. [38]

    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 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, Spri...

    2015

Pith tools

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