pith. sign in

arxiv: 2505.16786 · v2 · pith:TGSPSCNFnew · submitted 2025-05-22 · 💻 cs.LG

FlowMixer: A Depth-Agnostic Neural Architecture for Interpretable Spatiotemporal Forecasting

Fares B. Mehouachi , Saif Eddin Jabari This is my paper

Pith reviewed 2026-05-22 13:37 UTC · model grok-4.3

classification 💻 cs.LG
keywords spatiotemporal forecastingneural architectureinterpretable modelsKoopman theorysemi-group propertymatrix factorizationreversible networksdepth-agnostic learning
0
0 comments X

The pith

A single constrained layer with non-negative mixing and reversible maps can represent any network depth through composition while extracting interpretable Koopman eigenmodes for forecasting.

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

The paper presents FlowMixer as a depth-agnostic neural architecture for spatiotemporal forecasting. It builds the model from non-negative matrix mixing operations placed inside reversible mappings, which creates a semi-group structure. This structure lets one layer stand in for any number of layers by composition, removing the need to search over depths. The same design yields a Kronecker-Koopman eigenmodes representation that links the learned model to dynamical systems and permits algebraic adjustment of forecast horizons without retraining. Experiments across domains show that the resulting forecasts match state-of-the-art accuracy while supplying directly extractable, interpretable patterns for phenomena such as chaotic attractors and turbulent flows.

Core claim

FlowMixer shows that a single layer of constrained non-negative matrix mixing inside a reversible mapping framework possesses a semi-group property, allowing it to mathematically represent arbitrary depths through composition. The same construction produces a Kronecker-Koopman eigenmodes framework that extracts interpretable spatiotemporal patterns and supports direct algebraic manipulation of prediction horizons.

What carries the argument

The semi-group property generated by composing reversible non-negative matrix mixing operations, which encodes arbitrary depth and supplies the Kronecker-Koopman eigenmodes for interpretability.

If this is right

  • Depth search is eliminated for this class of spatiotemporal models.
  • Prediction horizons can be adjusted algebraically after training without retraining.
  • Eigenmodes extracted from the trained model provide direct links to dynamical systems analysis.
  • Competitive long-horizon forecasting accuracy is achieved while preserving mathematical interpretability across physical domains.

Where Pith is reading between the lines

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

  • Similar semi-group constructions could be tested in other sequence modeling tasks where depth selection currently dominates compute budgets.
  • The eigenmode representation may allow stability analysis or control design directly on the learned operator.
  • The reversible mixing framework might extend to settings that require invertible transformations for uncertainty quantification or data assimilation.

Load-bearing premise

Constrained non-negative matrix mixing layers inside reversible mappings can capture the full range of structured spatiotemporal dynamics without needing multiple layers or losing expressivity on complex physical phenomena.

What would settle it

A dataset of turbulent flow or chaotic attractors on which a multi-layer baseline consistently outperforms FlowMixer or on which the extracted eigenmodes fail to align with independently computed dynamical modes of the system.

Figures

Figures reproduced from arXiv: 2505.16786 by Fares B. Mehouachi, Saif Eddin Jabari.

Figure 1
Figure 1. Figure 1: Overview of FlowMixer’s constrained architecture. The core architecture (bottom) pro [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of Kronecker-Koopman Eigenmodes for traffic dataset. (a) Time eigenvalues [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Predictions of Lorenz (a), Rössler (b), and Aizawa (c) chaotic attractors (scaled [-1,1]). [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Prediction of vorticity fields for flow past a cylinder at [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FFT Spectral Analysis of Periodicity Patterns in Benchmark Time Series Datasets. The [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the FlowMixer architecture. Input data [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comprehensive ablation study of FlowMixer architectural components. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of forecasting performance for FlowMixer, NHiTS, and TSMixer on the [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FlowMixer trained for horizon h = 96 (green color), and later interpolated/extrapolated for various different horizons. Values of t closer to 1 yield best reslts naturally. Further constraints on the eigenvalues behaviour could improve the interpolation/extraploation capbilities and is a future direction of research [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: MSE for interpolated/extrapolated horizons. [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Predictions of Lorenz (a), Rössler (b), and Aizawa (c) chaotic attractors (scaled [-1,1]) [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparative analysis of Lorenz chaotic attractor predictions across three forecasting [PITH_FULL_IMAGE:figures/full_fig_p029_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison on Rossler attractor [PITH_FULL_IMAGE:figures/full_fig_p030_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison on Aizawa attractor 30 [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Eigenvalue analysis of the FlowMixer model for Lorenz, Rossler, and Aizawa attractors. [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Comparison of vorticity field predictions for 2D flow past a cylinder at Re=150. The figure [PITH_FULL_IMAGE:figures/full_fig_p034_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Comparative analysis of vorticity field predictions for 2D flow past a cylinder at Reynolds [PITH_FULL_IMAGE:figures/full_fig_p035_17.png] view at source ↗
read the original abstract

We introduce FlowMixer, a single-layer neural architecture that leverages constrained matrix operations to model structured spatiotemporal patterns with enhanced interpretability. FlowMixer incorporates non-negative matrix mixing layers within a reversible mapping framework - applying transforms before mixing and their inverses afterward. This shape-preserving design enables a Kronecker-Koopman eigenmodes framework that bridges statistical learning with dynamical systems theory, providing interpretable spatiotemporal patterns and facilitating direct algebraic manipulation of prediction horizons without retraining. The architecture's semi-group property enables this single layer to mathematically represent any depth through composition, eliminating depth search entirely. Extensive experiments across diverse domains demonstrate FlowMixer's long-horizon forecasting capabilities while effectively modeling physical phenomena such as chaotic attractors and turbulent flows. Our results achieve performance matching state-of-the-art methods while offering superior interpretability through directly extractable eigenmodes. This work suggests that architectural constraints can simultaneously maintain competitive performance and enhance mathematical interpretability in neural forecasting systems.

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.

Referee Report

2 major / 2 minor

Summary. The paper introduces FlowMixer, a single-layer neural architecture for spatiotemporal forecasting that employs non-negative matrix mixing layers inside a reversible mapping framework. It proposes a Kronecker-Koopman eigenmodes approach to bridge statistical learning and dynamical systems, enabling interpretable patterns and algebraic manipulation of prediction horizons. The central claim is that the architecture's semi-group property allows this single layer to represent arbitrary depths through composition, eliminating depth hyperparameter search. Experiments across domains are reported to match state-of-the-art performance while providing superior interpretability via directly extractable eigenmodes.

Significance. If the semi-group closure and expressivity preservation hold under the non-negativity and reversibility constraints, the work offers a mathematically grounded alternative to deep networks for forecasting, potentially removing depth tuning while linking to Koopman theory for interpretability. The direct horizon manipulation without retraining would be a practical strength if rigorously shown.

major comments (2)
  1. [Abstract and Section 3 (FlowMixer Architecture)] The manuscript's central claim (abstract and architecture section) that the semi-group property enables a single layer to mathematically represent any depth via composition rests on an unshown derivation of closure under composition for non-negative mixing matrices combined with reversible pre/post transforms. This is load-bearing for the depth-agnostic assertion, as non-negativity restricts mixing to conical combinations that may not preserve full expressivity for sign-changing or higher-order couplings.
  2. [Section 4 (Kronecker-Koopman Eigenmodes)] Section 4 (Kronecker-Koopman Eigenmodes): The eigenmodes are presented as bridging statistical learning and dynamical systems, but it is not shown whether they are derived independently of the fitted mixing parameters or extracted post-training from the learned matrices. If the latter, this risks circularity in the interpretability and algebraic manipulation claims.
minor comments (2)
  1. [Abstract] The abstract states performance matching state-of-the-art but provides no quantitative results, error bars, or ablation details on the depth-agnostic property; moving key metrics to the abstract would improve clarity.
  2. [Section 3] Notation for the constrained mixing matrix and reversible transforms could be standardized across equations to avoid ambiguity in the composition operation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment below and note the corresponding revisions.

read point-by-point responses

  1. Referee: [Abstract and Section 3 (FlowMixer Architecture)] The manuscript's central claim (abstract and architecture section) that the semi-group property enables a single layer to mathematically represent any depth via composition rests on an unshown derivation of closure under composition for non-negative mixing matrices combined with reversible pre/post transforms. This is load-bearing for the depth-agnostic assertion, as non-negativity restricts mixing to conical combinations that may not preserve full expressivity for sign-changing or higher-order couplings.

    Authors: We acknowledge that an explicit derivation of semi-group closure for the non-negative mixing matrices under the reversible pre/post-transform constraints was not provided in the original manuscript. In the revision we will insert a formal proof in Section 3 establishing that the constrained matrices remain closed under composition, thereby supporting the depth-agnostic claim. On expressivity, the reversible transforms expand the representable couplings beyond pure conical combinations; our experiments on chaotic and turbulent systems indicate that the resulting expressivity is adequate for the forecasting tasks considered. We will add a short discussion of this limitation and its empirical mitigation. revision: yes

  2. Referee: [Section 4 (Kronecker-Koopman Eigenmodes)] Section 4 (Kronecker-Koopman Eigenmodes): The eigenmodes are presented as bridging statistical learning and dynamical systems, but it is not shown whether they are derived independently of the fitted mixing parameters or extracted post-training from the learned matrices. If the latter, this risks circularity in the interpretability and algebraic manipulation claims.

    Authors: The eigenmodes are extracted after training from the learned mixing matrices via the Kronecker-Koopman construction in Section 4. This procedure is not circular: the parameters are optimized to minimize forecasting loss, after which the algebraic semi-group structure permits direct eigenmode manipulation for arbitrary horizons. We will revise Section 4 to state the extraction timing explicitly and to separate the training objective from the post-hoc algebraic interpretation. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims follow from architectural constraints without reduction to inputs by construction

full rationale

The paper defines FlowMixer via explicit design choices (non-negative matrix mixing inside reversible pre/post transforms) and then derives the semi-group property, Kronecker-Koopman eigenmodes, and depth-agnostic composition as consequences of those constraints. No equation or step is shown to define a quantity in terms of itself or to relabel a fitted parameter as an independent prediction. The algebraic horizon manipulation is presented as a direct result of the eigenmode structure extracted from the learned matrices, but this remains a post-design property rather than a tautological re-expression of the training data. The derivation chain is therefore self-contained against external benchmarks of expressivity and interpretability.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claims rest on the existence of a semi-group property for the reversible matrix operations and on the assumption that non-negative mixing preserves sufficient expressivity for chaotic and turbulent dynamics. No explicit free parameters are named, but the mixing matrices themselves are learned and therefore fitted. No new physical entities are postulated.

free parameters (1)
  • mixing matrix entries
    Non-negative matrix mixing layers are learned from data; their values are fitted parameters that determine the spatiotemporal patterns.
axioms (2)
  • domain assumption The composition of reversible transforms and matrix mixing forms a semi-group under which a single layer equals arbitrary depth.
    Invoked in the abstract to claim elimination of depth search; this mathematical property is asserted but not derived in the provided text.
  • domain assumption Kronecker-Koopman eigenmodes can be directly extracted from the learned architecture to yield interpretable spatiotemporal patterns.
    Central to the interpretability claim; assumes the framework bridges neural networks and dynamical systems without additional fitting steps.

pith-pipeline@v0.9.0 · 5694 in / 1505 out tokens · 30319 ms · 2026-05-22T13:37:18.640661+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

81 extracted references · 81 canonical work pages · 3 internal anchors

  1. [1]

    Model-free prediction of large spatiotemporally chaotic systems from data: A reservoir computing approach.Physical Review Letters, 120(2):024102, 2018

    Jaideep Pathak, Brian Hunt, Michelle Girvan, Zhixin Lu, and Edward Ott. Model-free prediction of large spatiotemporally chaotic systems from data: A reservoir computing approach.Physical Review Letters, 120(2):024102, 2018

    work page 2018

  2. [2]

    Enhancing computational fluid dynamics with machine learning.Nature Computational Science, 2(6):358–366, 2022

    Ricardo Vinuesa and Steven L Brunton. Enhancing computational fluid dynamics with machine learning.Nature Computational Science, 2(6):358–366, 2022

    work page 2022

  3. [3]

    Robust prediction of dynamical systems with structured neu- ral networks: Long-term behavior and chaos.Physica D: Nonlinear Phenomena, 427:133006, 2021

    Moonseop Han and Elizabeth Qian. Robust prediction of dynamical systems with structured neu- ral networks: Long-term behavior and chaos.Physica D: Nonlinear Phenomena, 427:133006, 2021

    work page 2021

  4. [4]

    John Wiley & Sons, Hoboken, New Jersey, 5 edition, 2015

    George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung.Time Series Analysis: Forecasting and Control. John Wiley & Sons, Hoboken, New Jersey, 5 edition, 2015

    work page 2015

  5. [5]

    Data-driven science and engineering: Machine learning, dynamical systems, and control.Cambridge University Press, 2019

    Steven L Brunton and J Nathan Kutz. Data-driven science and engineering: Machine learning, dynamical systems, and control.Cambridge University Press, 2019

    work page 2019

  6. [6]

    Promising directions of machine learning for partial differential equations.Nature Computational Science, 4(7):483–494, 2024

    Steven L Brunton and J Nathan Kutz. Promising directions of machine learning for partial differential equations.Nature Computational Science, 4(7):483–494, 2024

    work page 2024

  7. [7]

    Next generation reservoir computing.Nature communications, 12(1):1–8, 2021

    Daniel J Gauthier, Erik Bollt, Aaron Griffith, and Wendson AS Barbosa. Next generation reservoir computing.Nature communications, 12(1):1–8, 2021

    work page 2021

  8. [8]

    Higher-order granger reservoir computing: simultaneously achieving scalable complex structures inference and accurate dynamics prediction.Nature Communications, 15(1):2506, 2024

    Xin Li, Qunxi Zhu, Chengli Zhao, Xiaojun Duan, Bolin Zhao, Xue Zhang, Huanfei Ma, Jie Sun, and Wei Lin. Higher-order granger reservoir computing: simultaneously achieving scalable complex structures inference and accurate dynamics prediction.Nature Communications, 15(1):2506, 2024

    work page 2024

  9. [9]

    Reservoir computing using dynamic memristors for temporal information processing.Nature communica- tions, 8(1):2204, 2017

    Chao Du, Fuxi Cai, Mohammed A Zidan, Wen Ma, Seung Hwan Lee, and Wei D Lu. Reservoir computing using dynamic memristors for temporal information processing.Nature communica- tions, 8(1):2204, 2017

    work page 2017

  10. [10]

    Reservoir computing approaches for representation and classification of multivariate time series.IEEE transactions on neural networks and learning systems, 32(5):2169–2179, 2020

    Filippo Maria Bianchi, Simone Scardapane, Sigurd Løkse, and Robert Jenssen. Reservoir computing approaches for representation and classification of multivariate time series.IEEE transactions on neural networks and learning systems, 32(5):2169–2179, 2020

    work page 2020

  11. [11]

    Chaos as an interpretable benchmark for forecasting and data-driven modelling

    William Gilpin. Chaos as an interpretable benchmark for forecasting and data-driven modelling. arXiv preprint arXiv:2110.05266, 2021

    work page arXiv 2021

  12. [12]

    Backpropagation algorithms and reservoir computing in recurrent neural networks for the forecasting of complex spatiotemporal dynamics.Neural Networks, 126:191–217, 2020

    Pantelis R Vlachas, Jaideep Pathak, Brian R Hunt, Themistoklis P Sapsis, Michelle Girvan, Edward Ott, and Petros Koumoutsakos. Backpropagation algorithms and reservoir computing in recurrent neural networks for the forecasting of complex spatiotemporal dynamics.Neural Networks, 126:191–217, 2020

    work page 2020

  13. [13]

    Emerging opportuni- ties and challenges for the future of reservoir computing.Nature Communications, 15(1):2056, 2024

    Min Yan, Can Huang, Peter Bienstman, Peter Tino, Wei Lin, and Jie Sun. Emerging opportuni- ties and challenges for the future of reservoir computing.Nature Communications, 15(1):2056, 2024

    work page 2056

  14. [14]

    MIT Press, Cambridge, MA, 2012

    Masashi Sugiyama and Motoaki Kawanabe.Machine Learning in Non-Stationary Environments: Introduction to Covariate Shift Adaptation. MIT Press, Cambridge, MA, 2012

    work page 2012

  15. [15]

    Deterministic nonperiodic flow.Journal of atmospheric sciences, 20(2):130– 141, 1963

    Edward N Lorenz. Deterministic nonperiodic flow.Journal of atmospheric sciences, 20(2):130– 141, 1963

    work page 1963

  16. [16]

    An equation for continuous chaos.Physics Letters A, 57(5):397–398, 1976

    Otto E Rössler. An equation for continuous chaos.Physics Letters A, 57(5):397–398, 1976

    work page 1976

  17. [17]

    Chaos and fractals: Proceedings of the first ifip conference on fractals in the fundamental and applied sciences.North-Holland Delta Series, 1994

    Y Aizawa. Chaos and fractals: Proceedings of the first ifip conference on fractals in the fundamental and applied sciences.North-Holland Delta Series, 1994

    work page 1994

  18. [18]

    CRC Press, Boca Raton, 2 edition, 2018

    Steven H Strogatz.Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering. CRC Press, Boca Raton, 2 edition, 2018. 10

    work page 2018

  19. [19]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. InInternational Conference on Learning Representations, 2023

    work page 2023

  20. [20]

    Timemixer++: A general time series pattern machine for universal predictive analysis

    Shiyu Wang, Jiawei Li, Xiaoming Shi, Qinghe Liao, Mengan Liu, Qi Chen, Qingsong Wen, Zhiyuan Wang, Zijie Liu, Xiuwen Zhang, and Rong Jin. Timemixer++: A general time series pattern machine for universal predictive analysis. InarXiv preprint, 2024

    work page 2024

  21. [21]

    Wpmixer: Efficient multi-resolution mixing for long-term time series forecasting

    Md Mahmuddun Nabi Murad, Mehmet Aktukmak, and Yasin Yilmaz. Wpmixer: Efficient multi-resolution mixing for long-term time series forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, 2025

    work page 2025

  22. [22]

    Chimera: Effectively modeling multi- variate time series with 2-dimensional state space models

    Ali Behrouz, Michele Santacatterina, and Ramin Zabih. Chimera: Effectively modeling multi- variate time series with 2-dimensional state space models. InAdvances in Neural Information Processing Systems, 2024

    work page 2024

  23. [23]

    Model scale versus domain knowledge in statistical forecasting of chaotic systems.Physical Review Research, 5(4):043252, 2023

    William Gilpin. Model scale versus domain knowledge in statistical forecasting of chaotic systems.Physical Review Research, 5(4):043252, 2023

    work page 2023

  24. [24]

    Time-llm: Time series forecasting by reprogramming large language models

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Zhang, Pin Tang, Hao Cheng, Yong Luo, and Shuicheng Yan. Time-llm: Time series forecasting by reprogramming large language models. InInternational Conference on Learning Representations, 2024

    work page 2024

  25. [25]

    Zero-shot forecasting of chaotic systems.arXiv preprint arXiv:2409.15771, 2024

    Yuanzhao Zhang and William Gilpin. Zero-shot forecasting of chaotic systems.arXiv preprint arXiv:2409.15771, 2024

    work page arXiv 2024

  26. [26]

    Lee and H

    Daniel D. Lee and H. Sebastian Seung. Learning the parts of objects by non-negative matrix factorization.Nature, 401(6755):788–791, 1999

    work page 1999

  27. [27]

    Dynamic mode decomposition of numerical and experimental data.Journal of fluid mechanics, 656:5–28, 2010

    Peter J Schmid. Dynamic mode decomposition of numerical and experimental data.Journal of fluid mechanics, 656:5–28, 2010

    work page 2010

  28. [28]

    Arik, and Tomas Pfister

    Si-An Chen, Chun-Liang Li, Nathanael C Yoder, Sercan Ö Arık, and Tomas Pfister. Tsmixer: An all-mlp architecture for time series forecasting.arXiv preprint arXiv:2303.06053, 2023

    work page arXiv 2023

  29. [29]

    Reversible instance normalization for accurate time-series forecasting against distribution shift

    Taesu Kim, Jinhee Kim, Yunwon Seo, Jae Won Soh, and Jaegul Yoon. Reversible instance normalization for accurate time-series forecasting against distribution shift. InInternational Conference on Learning Representations, 2022

    work page 2022

  30. [30]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    work page 2017

  31. [31]

    Granger and Roselyne Joyeux

    Clive W.J. Granger and Roselyne Joyeux. An introduction to long-memory time series models and fractional differencing.Journal of Time Series Analysis, 1(1):15–29, 1980

    work page 1980

  32. [32]

    SIAM, Philadelphia, 2016

    J Nathan Kutz, Steven L Brunton, Bingni W Brunton, and Joshua L Proctor.Dynamic Mode Decomposition: Data-Driven Modeling of Complex Systems, volume 149 ofOther Titles in Applied Mathematics. SIAM, Philadelphia, 2016

    work page 2016

  33. [33]

    Ergodic theory, dynamic mode decomposition, and computation of spectral properties of the koopman operator.SIAM Journal on Applied Dynamical Systems, 16(4):2096–2126, 2017

    Hassan Arbabi and Igor Mezic. Ergodic theory, dynamic mode decomposition, and computation of spectral properties of the koopman operator.SIAM Journal on Applied Dynamical Systems, 16(4):2096–2126, 2017

    work page 2096

  34. [34]

    Chaos as an intermittently forced linear system.Nature communications, 8(1):19, 2017

    Steven L Brunton, Bingni W Brunton, Joshua L Proctor, Eurika Kaiser, and J Nathan Kutz. Chaos as an intermittently forced linear system.Nature communications, 8(1):19, 2017

    work page 2017

  35. [35]

    Hamiltonian systems and transformation in hilbert space.Proceedings of the National Academy of Sciences, 17(5):315–318, 1931

    Bernard O Koopman. Hamiltonian systems and transformation in hilbert space.Proceedings of the National Academy of Sciences, 17(5):315–318, 1931

    work page 1931

  36. [36]

    Brunton, J

    Steven L. Brunton, J. Nathan Kutz, and Joshua L. Proctor.Modern Koopman Theory for Dynamical Systems. Cambridge University Press, Cambridge, UK, 2022. 11

    work page 2022

  37. [37]

    Physics-informed dynamic mode decomposition.Proceedings of the Royal Society A, 479(2271):20220576, 2023

    Peter J Baddoo, Benjamin Herrmann, Beverley J McKeon, J Nathan Kutz, and Steven L Brunton. Physics-informed dynamic mode decomposition.Proceedings of the Royal Society A, 479(2271):20220576, 2023

    work page 2023

  38. [38]

    Modeling long-and short-term temporal patterns with deep neural networks

    Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling long-and short-term temporal patterns with deep neural networks. InThe 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, pages 95–104, 2018

    work page 2018

  39. [39]

    Mshyper: Multi-scale hyper- graph transformer for long-range time series forecasting

    Zongjiang Shang, Ling Chen, Binqing Wu, and Dongliang Cui. Mshyper: Multi-scale hyper- graph transformer for long-range time series forecasting. InarXiv preprint, 2024

    work page 2024

  40. [40]

    Ada-mshyper: Adaptive multi- scale hypergraph transformer for time series forecasting

    Zongjiang Shang, Ling Chen, Binqing Wu, and Dongliang Cui. Ada-mshyper: Adaptive multi- scale hypergraph transformer for time series forecasting. InAdvances in Neural Information Processing Systems, 2024

    work page 2024

  41. [41]

    Cyclenet: enhancing time series forecasting through modeling periodic patterns.Advances in Neural Information Processing Systems, 37:106315–106345, 2024

    Shengsheng Lin, Weiwei Lin, Xinyi Hu, Wentai Wu, Ruichao Mo, and Haocheng Zhong. Cyclenet: enhancing time series forecasting through modeling periodic patterns.Advances in Neural Information Processing Systems, 37:106315–106345, 2024

    work page 2024

  42. [42]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11106–11115, 2021

    work page 2021

  43. [43]

    Springer, Berlin, Heidelberg, 2013

    Martin Treiber and Arne Kesting.Traffic Flow Dynamics: Data, Models and Simulation. Springer, Berlin, Heidelberg, 2013

    work page 2013

  44. [44]

    Modeling and control of freeway traffic flow by state space neural networks.Neural Computing and Applications, 17(2):175–185, 2008

    Bart De Schutter and Serge P Hoogendoorn. Modeling and control of freeway traffic flow by state space neural networks.Neural Computing and Applications, 17(2):175–185, 2008

    work page 2008

  45. [45]

    Urban traffic analysis and forecasting through shared koopman eigenmodes.arXiv preprint arXiv:2409.04728, 2024

    Chuhan Yang, Fares B Mehouachi, Monica Menendez, and Saif Eddin Jabari. Urban traffic analysis and forecasting through shared koopman eigenmodes.arXiv preprint arXiv:2409.04728, 2024

    work page arXiv 2024

  46. [46]

    Forecasting of noisy chaotic systems with deep neural networks.Chaos, Solitons & Fractals, 153:111570, 2021

    Matteo Sangiorgio, Fabio Dercole, and Giorgio Guariso. Forecasting of noisy chaotic systems with deep neural networks.Chaos, Solitons & Fractals, 153:111570, 2021

    work page 2021

  47. [47]

    3d-integrated multilayered physical reservoir array for learning and forecasting time-series information.Nature Communications, 15(1):2044, 2024

    Sanghyeon Choi, Jaeho Shin, Gwanyeong Park, Jung Sun Eo, Jingon Jang, J Joshua Yang, and Gunuk Wang. 3d-integrated multilayered physical reservoir array for learning and forecasting time-series information.Nature Communications, 15(1):2044, 2024

    work page 2044

  48. [48]

    Deep learning.Nature, 521(7553):436–444, 2015

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.Nature, 521(7553):436–444, 2015

    work page 2015

  49. [49]

    Robust forecasting using predictive generalized synchronization in reservoir computing.Chaos: An Interdisciplinary Journal of Nonlinear Science, 31(12):123118, 2021

    Jason A Platt, Alexander S Wong, Ryan Clark, Stephen G Penny, and Henry DI Abarbanel. Robust forecasting using predictive generalized synchronization in reservoir computing.Chaos: An Interdisciplinary Journal of Nonlinear Science, 31(12):123118, 2021

    work page 2021

  50. [50]

    Data-driven prediction of chaotic dynamics: Methods, challenges, and opportunities.Annual Review of Condensed Matter Physics, 14:379–401, 2023

    Jaideep Pathak, Brian R Hunt, Georg M Goerg, and Themistoklis P Sapsis. Data-driven prediction of chaotic dynamics: Methods, challenges, and opportunities.Annual Review of Condensed Matter Physics, 14:379–401, 2023

    work page 2023

  51. [51]

    N-beats: Neural basis expansion analysis for interpretable time series forecasting.arXiv preprint arXiv:2002.05746, 2020

    Boris N Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio. N-beats: Neural basis expansion analysis for interpretable time series forecasting.arXiv preprint arXiv:2002.05746, 2020

    work page arXiv 2002

  52. [52]

    Machine learning for fluid mechanics.Annual Review of Fluid Mechanics, 52:477–508, 2020

    Steven L Brunton, Bernd R Noack, and Petros Koumoutsakos. Machine learning for fluid mechanics.Annual Review of Fluid Mechanics, 52:477–508, 2020

    work page 2020

  53. [53]

    Maziar Raissi, Paris Perdikaris, and George 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, 378:686–707, 2019

    work page 2019

  54. [54]

    Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021. 12

    work page 2021

  55. [55]

    When and why pinns fail to train: A neural tangent kernel perspective.Journal of Computational Physics, 449:110768, 2022

    Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective.Journal of Computational Physics, 449:110768, 2022

    work page 2022

  56. [56]

    Hidden physics models: Machine learning of nonlinear partial differential equations.Journal of Computational Physics, 357:125–141, 2020

    Maziar Raissi and George Em Karniadakis. Hidden physics models: Machine learning of nonlinear partial differential equations.Journal of Computational Physics, 357:125–141, 2020

    work page 2020

  57. [57]

    Towards physics- informed deep learning for turbulent flow prediction

    Rui Wang, Karthik Kashinath, Mustafa Mustafa, Adrian Albert, and Rose Yu. Towards physics- informed deep learning for turbulent flow prediction. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1457–1466. ACM, 2020

    work page 2020

  58. [58]

    Coarse-grained physics-based predic- tion of three-dimensional unsteady flows via neural networks.Science Advances, 7(46):eabj0751, 2021

    Alejandro Güemes, Stefano Discetti, and Andrea Ianiro. Coarse-grained physics-based predic- tion of three-dimensional unsteady flows via neural networks.Science Advances, 7(46):eabj0751, 2021

    work page 2021

  59. [59]

    Embedding hard physical constraints in neural network coarse-graining of 3d turbulence.arXiv preprint arXiv:2002.00021, 2020

    Arvind T Mohan, Nicholas Lubbers, Daniel Livescu, and Michael Chertkov. Embedding hard physical constraints in neural network coarse-graining of 3d turbulence.arXiv preprint arXiv:2002.00021, 2020

    work page arXiv 2002

  60. [60]

    Learning the solution operator of para- metric partial differential equations with physics-informed deeponets.Science Advances, 7(40):eabi8605, 2021

    Sifan Wang, Shyam Sankaran, and Paris Perdikaris. Learning the solution operator of para- metric partial differential equations with physics-informed deeponets.Science Advances, 7(40):eabi8605, 2021

    work page 2021

  61. [61]

    Enforcing analytic constraints in neural networks emulating physical systems.Physical Review Letters, 126(9):098302, 2021

    Tom Beucler, Michael Pritchard, Stephan Rasp, Jordan Ott, Pierre Baldi, and Pierre Gentine. Enforcing analytic constraints in neural networks emulating physical systems.Physical Review Letters, 126(9):098302, 2021

    work page 2021

  62. [62]

    N-hits: Neural hierarchical interpolation for time series forecasting

    Cristian Challu, Kin G Olivares, Boris N Oreshkin, Federico Garza, Max Mergenthaler-Canseco, and Artur Dubrawski. N-hits: Neural hierarchical interpolation for time series forecasting. arXiv preprint arXiv:2201.12886, 2022

    work page arXiv 2022

  63. [63]

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting.In International Conference on Machine Learning, pages 26501–26518, 2022

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting.In International Conference on Machine Learning, pages 26501–26518, 2022

    work page 2022

  64. [64]

    Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. InAdvances in Neural Information Processing Systems, volume 34, pages 22419–22430, 2021

    work page 2021

  65. [65]

    Are transformers effective for time series forecasting? InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023

    work page 2023

  66. [66]

    Koopa: Learning non-stationary time series dynamics with koopman predictors.Advances in neural information processing systems, 36:12271–12290, 2023

    Yong Liu, Chenyu Li, Jianmin Wang, and Mingsheng Long. Koopa: Learning non-stationary time series dynamics with koopman predictors.Advances in neural information processing systems, 36:12271–12290, 2023

    work page 2023

  67. [67]

    Koopman neural forecaster for time-series with temporal distribution shifts.Proceedings of ICLR, 2, 2023

    Qi Rose Yu, Ray Wang, Sercan Arik, and Yihe Dong. Koopman neural forecaster for time-series with temporal distribution shifts.Proceedings of ICLR, 2, 2023

    work page 2023

  68. [68]

    Exact stochastic newton method for deep learning: the feedforward networks case

    Fares B Mehouachi and Chaouki Kasmi. Exact stochastic newton method for deep learning: the feedforward networks case. 2022

    work page 2022

  69. [69]

    Exact stochastic second order deep learning.arXiv preprint arXiv:2104.03804, 2021

    Fares B Mehouachi and Chaouki Kasmi. Exact stochastic second order deep learning.arXiv preprint arXiv:2104.03804, 2021

    work page arXiv 2021

  70. [70]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differen- tial equations.arXiv preprint arXiv:2010.08895, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  71. [71]

    Detection of uavs based on spectrum monitoring and deep learning in negative snr conditions.URSI Radio Sci

    Fares B Mehouachi, Qingjie Yang, Juan Galvis, Santiago Morales, Milosch Meriac, Felix Vega, and Chaouki Kasmi. Detection of uavs based on spectrum monitoring and deep learning in negative snr conditions.URSI Radio Sci. Lett., 3:43, 2021. 13

    work page 2021

  72. [72]

    Detection and identification of uavs based on spectrum monitoring and deep learning in negative snr conditions.URSI GASS, 2021

    F Mehouachi, Juan Galvis, Santiago Morales, Milosch Meriac, Felix Vega, and Chaouki Kasmi. Detection and identification of uavs based on spectrum monitoring and deep learning in negative snr conditions.URSI GASS, 2021

    work page 2021

  73. [73]

    Catastrophic Overfitting, Entropy Gap and Participation Ratio: A Noiseless $l^p$ Norm Solution for Fast Adversarial Training

    Fares B Mehouachi and Saif Eddin Jabari. Catastrophic overfitting, entropy gap and participation ratio: A noiseless lp norm solution for fast adversarial training.arXiv preprint arXiv:2505.02360, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  74. [74]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInterna- tional Conference on Learning Representations, 2014

    work page 2014

  75. [75]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam.arXiv preprint arXiv:1711.05101, 2017. 14 Impact Statement FlowMixer introduces an interpretable approach to spatiotemporal forecasting that bridges statistical learning with dynamical systems theory. By employing constrained matrix operations within a reversible mapping framewor...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  76. [76]

    Automatic Parameter Tuning: FlowMixer learns the optimal α directly from data rather than requiring a priori estimation

    work page

  77. [77]

    24 J Chaotic Systems Prediction J.1 Overview We evaluated FlowMixer on three canonical chaotic systems with distinct dynamical characteristics

    Computational Tractability: The first-order approximation maintains efficiency while capturing essential long-memory characteristics Through this mechanism, FlowMixer effectively addresses non-stationarity through a theoretically grounded, differentiable approach inspired by classical time series models. 24 J Chaotic Systems Prediction J.1 Overview We eva...

    work page 2018

  78. [78]

    We first compute an intermediate velocity fieldu ∗ without the pressure gradient: u∗ −u n ∆t =−(u n · ∇)un + 1 Re ∇2un

    work page

  79. [79]

    The viscous terms are treated using an ADI scheme to ensure stability:    (I− ∆t 2Re ∂2 ∂x2 )ˆu=u n + ∆t 2Re ∂2un ∂y 2 (I− ∆t 2Re ∂2 ∂y 2 )u∗ = ˆu+ ∆t 2Re ∂2ˆu ∂x2 (29)

    work page

  80. [80]

    The pressure Poisson equation is then solved to enforce incompressibility: ∇2pn+1 = 1 ∆t ∇ ·u ∗ (30)

    work page

Showing first 80 references.