WorksPhysics and Astronomy
A catalog of exact steady states in Kolmogorov flow
Abstract
Two-dimensional Kolmogorov flow is a common proving ground for the dynamical systems picture of turbulence. Hundreds of unstable periodic orbits have been found for this system across many Reynolds numbers and domain sizes, although the most well explored is Re on a square domain with a forcing wavenumber of Here I present a catalog of 8952 exact steady states of the flow in this canonical parameter regime. These solutions provide a fixed constellation informing the dynamics of both turbulence and transients.
Introduction
It has been a community effort over recent decades to realize the dynamical systems picture of turbulence in which the geometry and topology of turbulence is determined by invariant sets: equilibria, periodic orbits, higher dimensional tori, and the stable and unstable manifolds of these solutions. These exact solutions to the Navier-Stokes equations have been called Exact Coherent Structures (ECS) because they are exact solutions to the hydrodynamic model and have historically been used to understand the recurrence of large coherent structures observed in turbulence. This picture can be attributed in part to Eberhard Hopf (Hopf 1948) who viewed turbulence as a finite dimensional manifold in the space of admissible flow fields, although we know now that chaotic dynamics tend to produce fractal chaotic sets rather than smooth manifolds [10].
The primary barrier to the dynamical systems picture is obvious to any practitioner: the numerical computation of ECS is difficult. Any invariant set requires optimization over a high dimensional space requiring significant memory for variational methods [18, 25] or significant walltime for shooting methods [4]. While steady state solutions are the cheapest to converge, periodic orbits provide the dynamical skeleton of turbulence [7]. Periodic Orbit Theory (POT) predicts that ergodic measures of strange attractors can be expressed
Appendix A
as weighted sums of unstable periodic orbits that densely cover the attractor (Cvitanović et al. [8]).
Kolmogorov flow was suggested by Andrey Kolmogorov in 1959 (Arnold & Meshalkin [1]). A fluid with periodic boundary conditions is forced monochromatically to drive various cascades. 2D Kolmogorov flow has been of wide interest to dynamical systems researchers given its dynamical complexity and simple implementation ([27, 17, 4, 32, 20, 21, 12, 33, 30, 14, 31, 15, 29, 25, 22, 9, 6, 34, 24, 11, 5, 35, 26, 36, 2] (Farazmand & Sapsis 2016)).1
A common theme of these investigations is the convergence of periodic orbits; these solutions shadow turbulence and give insight into the underpinning nonlinear processes of the flow. At this time, hundreds of periodic orbits are known by several active research groups. These are usually identified by seeding a Newton-Raphson algorithm with near recurrences from turbulence. While these loops provide a dynamical skeleton of the chaotic set, the simpler set of steady states provides a constellation. These steady states are fixed points, and they constrain the topology and geometry of the flow. Three such solutions were provided in Chandler & Kerswell [4] and twenty-five were later provided in Farazmand [12]. Here I provide a constellation orders of magnitude larger: 8952 steady states identified in 2D Kolmogorov flow at .
Numerical methods
Kolmogorov flow is a choice of geometry and forcing for the incompressible Navier-Stokes equation
I will work with , with periodic boundary conditions, and following Chandler & Kerswell [4]. It is advantageous to work in the vorticity-velocity formulation defining to obtain the pressure-free vorticity equation
The vorticity field is stored numerically, and the velocity field is obtained by solving the simple linear system and in Fourier space. The mean flow is undetermined by these constraints, so I enforce at all times, where
The vorticity forcing has continuous translation symmetry in , so this system supports steady state traveling waves satisfying . Combining this wave equation and the vorticity equation gives the spatial constraint
Note that equilibria are just a special case of traveling wave with . This equation is the primary concern of this manuscript. I discretize this equation pseudospectrally on a grid with Fast Fourier Transforms (FFTs). The derivatives and velocity evaluation are done in Fourier space with the quadratic nonlinearity being computed by pointwise multiplication followed by 2/3rds dealiasing ([23]) to preserve exact continuous translational symmetry.
An initial guess for a pair solving Equation (5) is converged with Newton-GMRES. A phase condition is added to the Jacobian to prevent motion along the trivial marginal direction . A preconditioner of was found critical to the success of GMRES ([28]). The Newton step is obtained by solving
This search deviates from the literature in two important ways.
Initial guesses for traveling waves were never seeded from turbulence. Instead, initial guesses were Fourier filtered random noise with zero mean rescaled to have a standard deviation of ; the wave speed was drawn from a normal distribution with zero mean and . The cutoff in Fourier space was varied from a maximum wavenumber of 4 to 6. All initial guess parameters were dynamically updated by Claude to maximize the probability of success.
The Newton-GMRES procedure was designed to be lazy. Generating a new initial guess is effectively free, so I encouraged abandoning Newton iteration if solutions did not show signs of rapid convergence. This is akin to buying a new lottery ticket as soon as your current ticket seems unlikely to win. This attitude was inspired by conversations at the long program Multi-Fidelity Methods for Fusion Energy hosted by the Institute for Pure and Applied Mathematics (IPAM). The initial guesses are converged in single precision (float32) with an adaptive GMRES tolerance of clamped between 0.01 and 0.5. If the solution did not reach in 32 Newton-GMRES iterations, then the initial condition is declared a failure and discarded. If single precision iteration does reach this target, then a double precision (float64) Newton-GMRES is kicked off with a fixed GMRES tolerance of . If the state reaches a maximum pointwise error of less than in 80 Newton steps, it is accepted as a candidate solution.
This Newton-GMRES pipeline for Kolmogorov flow was entirely written with Claude Code in C, but the discovered solutions were carefully hand-validated in MATLAB and Python. While Claude rapidly produced source code, there was a nontrivial debugging stage: I had to confront Claude over inconsistent ordering, incorrect dealiasing, and failure to set the mean vorticity to zero as required by periodic boundary conditions. While Claude made this project possible and more enjoyable, it is my opinion that current LLMs are not capable of fully autonomous research.
The resulting C code was ran in parallel on 8 cores for three months on an ASUS Expertbook with an Intel Core Ultra 5. This pipeline was migrated to an Ampere Altra Q64 and ran on all 64 cores for an additional month. No GPU was ever used to converge new solutions, although an NVIDIA V100 computed the linear stability of all solutions. I am choosing to halt the search for steady states now as it is increasingly impractical to converge new solutions.
Results
The final catalog contains a total of 8952 steady state solutions comprised of 847 equilibria () and 8105 traveling waves (). Solutions are considered distinct if after optimizing over symmetries, the maximum pointwise difference in vorticity is greater than . There are many solutions near bifurcations, so it is likely that there are true solutions that fall below this discrimination threshold and were ignored as duplicates.
These solutions were clustered with recursive application of the HDBSCAN algorithm (Campello et al. [3]) implemented in scikit-learn. The metric used for clustering is the symmetry optimized distance between solutions. Both the 16 discrete symmetries (Chandler & Kerswell [4]) and continuous translation are optimized over. This distance is a true metric so the optimized distance satisfies the triangle inequality. I use cluster_selection_method=”leaf” rather than the scikit-learn default ”eom”, which favors merging into a small number of large clusters. Because a single HDBSCAN pass leaves many solutions without a cluster assignment, I recursively re-cluster whatever remains unassigned at each level until a pass finds no further structure, which occurred after 39 levels. This procedure yields 416 families, containing 8294 of the 8952 solutions (92.6%); the remaining 658 (7.3%) are not assigned to any cluster. I do not force these into the nearest family: HDBSCAN is a density-based method for which leaving outliers unclassified is expected behavior, not a shortcoming. The single largest family (628 members) does not itself subdivide further under the same procedure at any tested threshold.
This clustering is not intended to be definitive. Rather, it is a guide for interactive exploration of these solutions. Solutions are named according to their family. For example, the seventh member of family 101 is named F101-7. The solutions which could not be clustered instead have the prefix N for “noise”. Families are ordered by the mean number of unstable directions. Family 1 has the smallest number of unstable directions on average. The ordering within a family is the order in which solutions were discovered by the search.
Figure 1 shows a 2D projection of all solutions. This projection uses the physical quantities of kinetic energy and dissipation
Figure 1. Low dimensional projection of all steady state solutions. All 416 families are shown in different colors, with the noise solutions in gray. Family 1 can be seen as the low dissipation cluster in blue.
Figure 1 shows that the majority of families cannot be meaningfully distinguished by alone. An outlier is Family 1, which is shown as a low dissipation cluster in blue. This family is relevant to turbulence, as seen in Figure 2.
Figure 2. A histogram of dissipation for the steady state solutions in the catalog and a turbulent trajectory. Note that the catalog extends to dissipations that are larger than seen in the chaotic set.
Figure 2 shows a histogram of dissipation for both the catalog of steady states and a turbulent trajectory. Clearly the catalog extends to dissipation values far larger than typically seen in the chaotic set. Such solutions are relevant for transients, but not late time turbulence. Further, there is a clear gap in for steady state solutions. This gap is physical and is where turbulence spends most of its time. This region of state space is filled with periodic orbits.
Wave speeds were found over the range . A histogram of these speeds is shown in Figure 3. Note that there is a clear separation between traveling waves and equilibria. The distribution of does not smoothly approach . Further, the distribution of appears to be bimodal. A small population of high speed solutions is visible. Figure 4 shows how these wave speeds correlate with dissipation. Low solutions have no significant dependence on ; however, the high solutions are associated with high .
Figure 3. A histogram of wave speeds. A factor of is added so that equilibria () can appear.
Figure 4. Low dimensional projection of dissipation and wave speed . The cluster of high wave speed solutions is associated with higher dissipation. Color corresponds to family. Again family 1 is visible in blue as low dissipation outliers.
Linear stability
For every one of the 8952 solutions I computed the leading 32 eigenvalues and real Schur vectors of the operator obtained by integrating the linearized vorticity equation to . This integration was performed with Lawson-RK4 [19] in the solution’s comoving frame. If the 32nd vector corresponded to an unresolved complex conjugate pair, then it was discarded. If the number of unstable directions was larger than 32, a larger set of 64 eigenvectors was converged. All solutions had fewer than 64 unstable modes. The real Schur decomposition of this operator was estimated with a restarted block Arnoldi–Krylov–Schur method (block size 40, 8 blocks per cycle, up to 10 restart cycles, Ritz-pair residual tolerance ). Figure 5 shows a histogram of unstable directions with a minimum of 2 and maximum of 42 with a mean of 22.0087.
Figure 5. A histogram of unstable directions for all solutions in the catalog. No solution was found to be linearly stable.
Finite time Floquet multipliers were computed, so only the real part of the continuous time eigenvalues can be estimated reliably with . The maximum growth rate is likely dominant under a random perturbation. growth rate is shown against dissipation in Figure 6. There is a cluster of highly unstable solutions corresponding to higher dissipation.
Figure 6. Leading linear growth rate versus dissipation. The cluster of highly unstable solutions is associated with higher dissipation. Again Family 1 is visible as a low dissipation outlier in blue.
Dynamical relevance to turbulence
A turbulent trajectory was simulated for timesteps with . Each point was compared to the catalog to find the nearest steady state after optimizing over all 16 discrete symmetries and continuous translation. To accomplish this, the first solution in each family was used to create a pivot table: a matrix of distances between pivots and a matrix of distances between pivots and the entire catalog. Since the symmetry optimized distance satisfies the triangle inequality, I can quickly determine the closest steady state to a turbulent snapshot by only optimizing distance for states that the triangle inequality does not eliminate.
Figure 7(a) shows the normalized distance from turbulence to the nearest steady state in the catalog. The color corresponds to the family of the nearest steady state. At low dissipation, Family 1 dominates. Other families only become relevant when .
Figure 7. (a) distance to the closest steady state for a turbulent trajectory. This distance is normalized by the norm of solution F1-1. Each color corresponds to a different family, with Family 1 being the large blue core at low dissipation. (b) Histogram of nearest family to a turbulent state. Note that Family 1 is most relevant by an order of magnitude. More unstable families are visited less frequently, with Families 371-416 not visited at all by this turbulent segment. More than 90% of turbulent states are closest to Family 1.
Figure 7(b) shows the frequency of visiting each family. Family 1 dominates with 90.26% of turbulent states being closest to this low dissipation cluster. This is unsurprising, as Family 1 is both low in unstable directions and in max (see Figure 6).
Figure 7(b) may give the impression that a sparse subset of families are closest to turbulence. This used a turbulent segment of length . The rarity of extreme high dissipation events makes it difficult to quantify which families are truly relevant to extreme behavior. A much longer turbulent segment would be required to make such statements. Rather, Figure 7(b) suggests that a wide range of families are relevant to extreme events. However, the distance to steady states during extreme events can be times larger than the distance during the shadowing of Family 1.
Discussion and conclusion
Running an LLM-driven Newton–GMRES search produced a vast catalog of 8952 steady state solutions of 2D Kolmogorov flow. These solutions have been clustered into 416 families using a symmetry optimized distance. No forward time integration was used to converge these solutions, so they span both the dissipation regimes of turbulence and bizarre high dissipation regimes only seen in initial transients. These steady states form a fixed constellation rather than a dynamical skeleton of turbulence, but they are valuable landmarks for anyone interested in the dynamical systems picture of Kolmogorov flow.
This work would not have been possible without Claude Code. The entire pipeline was written in C by agents and carefully debugged and validated by me. The agents had several bugs and struggled with maintaining conventions across the codebase. They were not capable of fully autonomous research, but could achieve high quality solutions with guidance.
Data availability statement. The full solution database solutions.h5 (1.09GB) is available at https://huggingface.co/datasets/mgolden30/Kolmogorov_TW. This dataset along with the linear stability (38GB) is available at https://invariant-sets.org.
Acknowledgements. I am grateful to the UCLA Institute for Pure and Applied Mathematics (IPAM), which allowed me to discuss this work with George Holt, Phil Travis, Clarence Rowley, and many others. I am grateful for helpful discussions with Predrag Cvitanović, Roman Grigoriev, Mike Schatz, and Dmitriy Zhigunov.
Funding. Part of this research was performed while the author was visiting the Institute for Pure and Applied Mathematics (IPAM), which is supported by the National Science Foundation (Grant No. DMS-1925919/2422832).
Declaration of interests. The author reports no conflict of interest.
Representative samples
Visualizing all 8952 solutions in a PDF document is unwise. Here I provide a random set of 400 solutions in Figure 8 and a complete visualization of Family 1 in Figure 9. All solutions can be interacted with at https://invariant-sets.org.
Figure 8. A random selection of 400 steady state solutions. Vorticity is shown with color limits .
Figure 9. The 24 solutions making up Family 1.
References
- [1]Arnold, V.I & Meshalkin, LD 1959 The seminar on a kolmogorov on selected topics in analysis. Usp. Mat. Nauk 15, 247–250.
- [2]Beck, Pierre & Schneider, Tobias M 2026 Identifying recurrent flows in high-dimensional dissipative chaos from low-dimensional embeddings. arXiv preprint arXiv:2601.0590.arxiv.org/abs/2601.01590
- [3]Campello, Ricardo JGB, Moulavi, Davoud & Sander, Jörg 2013 Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pp. 160–172. Springer.DOI
- [4]Chandler, Gary J & Kerswell, Rich R 2013 Invariant recurrent solutions embedded in a turbulent two-dimensional kolmogorov flow. Journal of Fluid Mechanics 722, 554–595.DOI
- [5]Cleary, Andrew & Page, Jacob 2025 Dynamical recurrence of periodic orbits under increasing reynolds number and connections to inviscid dynamics. Journal of Fluid Mechanics 1020, A52.DOI
- [6]Constante-Amores, C Ricardo & Graham, Michael D 2024 Data-driven state-space and koopman operator models of coherent state dynamics on invariant manifolds. Journal of Fluid Mechanics 984, R9.DOI
- [7]Cvitanović, Predrag 1991 Periodic orbits as the skeleton of classical and quantum chaos. Physica D: Nonlinear Phenomena 51 (1–3), 138–151.DOI
- [8]Cvitanović, Predrag, Artuso, Roberto, Mainieri, Renato, Tanner, Gregg & Vattay, Gábor 2020 Chaos: Classical and Quantum. Niels Bohr Institute.
- [9]De Jesús, Carlos E Pérez & Graham, Michael D 2023 Data-driven low-dimensional dynamic model of kolmogorov flow. Physical Review Fluids 8 (4), 044402.
- [10]Eckmann, J-P & Ruelle, David 1985 Ergodic theory of chaos and strange attractors. Reviews of modern physics 57 (3), 617.
- [11]Engel, Matthias, Ashtari, Omid, Schneider, Tobias M & Linkmann, Moritz 2025 Search for unstable relative periodic orbits in channel flow using symmetry-reduced dynamic mode decomposition. Journal of Fluid Mechanics 1013, A45.DOI
- [12]Farazmand, Mohammad 2016 An adjoint-based approach for finding invariant solutions of navier–stokes equations. Journal of Fluid Mechanics 795, 278–312.DOI
- [14]Farazmand, Mohammad & Sapsis, Themistoklis P 2017 A variational approach to probing extreme events in turbulent dynamical systems. Science advances 3 (9), e1701533.DOI
- [15]Farazmand, Mohammad & Sapsis, Themistoklis P 2019 Closed-loop adaptive control of extreme events in a turbulent flow. Physical Review E 100 (3), 033110.DOI
- [17]Inubushi, Masanobu, Kobayashi, Miki U, Takehiro, Shin-ichi & Yamada, Michio 2012 Covariant lyapunov analysis of chaotic kolmogorov flows. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics 85 (1), 016331.pubmed.ncbi.nlm.nih.gov/22400681
- [18]Lan, Yueheng & Cvitanović, Predrag 2004 Variational method for finding periodic orbits in a general flow. Physical Review E 69 (1), 016217.pubmed.ncbi.nlm.nih.gov/14995703
- [19]Lawson, J Douglas 1967 Generalized runge-kutta processes for stable systems with large lipschitz constants. SIAM Journal on Numerical Analysis 4 (3), 372–380.DOI
- [20]Lucas, Dan & Kerswell, Rich 2014 Spatiotemporal dynamics in two-dimensional kolmogorov flow over large domains. Journal of fluid mechanics 750, 518–554.
- [21]Lucas, Dan & Kerswell, Rich R 2015 Recurrent flow analysis in spatiotemporally chaotic 2-dimensional kolmogorov flow. Physics of Fluids 27 (4).
- [22]Lucas, Dan & Yasuda, Tatsuya 2022 Stabilization of exact coherent structures in two-dimensional turbulence using time-delayed feedback. Physical Review Fluids 7 (1), 014401.
- [23]Orszag, Steven A 1971 On the elimination of aliasing in finite-difference schemes by filtering high-wavenumber components. Journal of Atmospheric Sciences 28 (6), 1074–1074.doi.org/10.1175/1520-0469(1971)028<1074:oteoai>2.0.co;2
- [24]Page, Jacob 2025 Computation of simple invariant solutions in fluid turbulence with the aid of deep learning: J. page. Nonlinear dynamics 113 (23), 31895–31915.DOI
- [25]Parker, Jeremy P & Schneider, Tobias M 2022 Variational methods for finding periodic orbits in the incompressible navier–stokes equations. Journal of Fluid Mechanics 941, A17.DOI
- [26]Parker, Jeremy P & Schneider, Tobias M 2026 From synthetic turbulence to true solutions: A deep diffusion model for discovering periodic orbits in the navier-stokes equations. arXiv preprint arXiv:2602.23181.
- [27]Platt, Nathan, Sirovich, Lawrence & Fitzmaurice, Nessan 1991 An investigation of chaotic kolmogorov flows. Physics of Fluids A: Fluid Dynamics 3 (4), 681–696.
- [28]Saad, Youcef & Schultz, Martin H 1986 Gmres: A generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM Journal on scientific and statistical computing 7 (3), 856–869.DOI
- [29]Suri, Balachandra, Pallantla, Ravi Kumar, Schatz, Michael F & Grigoriev, Roman O 2019 Heteroclinic and homoclinic connections in a kolmogorov-like flow. Physical Review E 100 (1), 013112.DOI
- [30]Suri, Balachandra, Tithof, Jeffrey, Grigoriev, Roman O & Schatz, Michael F 2017 Forecasting fluid flows using the geometry of turbulence. Physical Review letters 118 (11), 114501.DOI
- [31]Suri, Balachandra, Tithof, Jeffrey, Grigoriev, Roman O & Schatz, Michael F 2018 Unstable equilibria and invariant manifolds in quasi-two-dimensional kolmogorov-like flow. Physical Review E 98 (2), 023105.DOI
- [32]Suri, Balachandra, Tithof, Jeffrey, Mitchell, Radford, Grigoriev, Roman O & Schatz, Michael F 2014 Velocity profile in a two-layer kolmogorov-like flow. Physics of Fluids 26 (5).DOI
- [33]Tithof, Jeffrey, Suri, Balachandra, Pallantla, Ravi Kumar, Grigoriev, Roman O & Schatz, Michael F 2017 Bifurcations in a quasi-two-dimensional kolmogorov-like flow. Journal of Fluid Mechanics 828, 837–866.DOI
- [34]Vela-Martín, Alberto & Avila, Marc 2024 Large-scale patterns set the predictability limit of extreme events in kolmogorov flow. Journal of Fluid Mechanics 986, A2.
- [35]VINOGRAD, MELISA Y, CULLEN, JOAQUÍN & CLARK DI LEONI, PATRICIO 2026 Dimensional regimes in kolmogorov flow. Physical Review Fluids 11 (3), 034402.
- [36]ZHIGUNOV, DMITRY & PAGE, JACOB 2026 Exact coherent structures as building blocks of turbulence on large domains. arXiv preprint arXiv:2601.21970 .arxiv.org/abs/2601.21970