Unsupervised learning for anticipating critical transitions

Bryan Glaz; Ling-Wei Kong; Mulugeta Haile; Shirin Panahi; Ying-Cheng Lai

REVIEW 4 major objections 6 minor 1 cited by

Unsupervised learning for anticipating critical transitions

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a variational autoencoder can extract a system's hidden bifurcation parameter from time series alone, and that feeding that extracted parameter into a parameter-driven reservoir computer lets it anticipate critical…

desk verdict Useful combination of VAE and parameter-driven reservoir computing for anticipating critical transitions, but the load-bearing extrapolation is not separated from calibration, so the headline claim is only partially supported. read the letter →

arxiv 2501.01579 v1 pith:SZEEZKB4 submitted 2025-01-02 nlin.CD cs.LG

Shirin Panahi , Ling-Wei Kong , Bryan Glaz , Mulugeta Haile , Ying-Cheng Lai This is my paper · ORCID

classification nlin.CDcs.LG MSC 37M1037G1068T07 PACS 05.45.-a

keywords criticaltransitionsbifurcationparameterextractionvariationalautoencoderreservoircomputingunsupervisedlearningtippingpointpredictionKuramoto-SivashinskyequationLorenzsystem

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

The paper aims to establish that a machine-learning pipeline can anticipate a critical transition of a complex dynamical system without ever being told the system's bifurcation parameter. A variational autoencoder (VAE) is trained on time series from the normal regime, and one of its latent variables takes on the role of the hidden driving parameter; that extracted parameter is then fed into a parameter-driven reservoir computer. The authors demonstrate the full pipeline on the Lorenz system with one and two parameters, on the spatiotemporal Kuramoto-Sivashinsky equation, on partial state observations, and on a three-species food chain, reporting accurate predictions of the transition point in each case. The practical payoff, if the claim holds, is that tipping-point early warning no longer requires measuring or knowing the control parameter that is about to trigger collapse.

What carries the argument

The central object is the latent variable $z$ at the VAE bottleneck, treated as a surrogate for the unknown bifurcation parameter. The encoder is a deep convolutional neural network that maps each input time series to a Gaussian distribution with mean $\mu_z$ and variance $\sigma_z^2$; the decoder is a feedforward network that propagates an initial condition forward in time, modulated by $z$. Training minimizes a reconstruction loss plus a three-part regularization $R_z = E_1 + E_2 + E_3$, where $E_1$ controls mutual information between latent variables and data, $E_2$ penalizes redundancy among latent channels, and $E_3$ pulls each channel toward a unit Gaussian prior, so that only informative channels escape collapse. The surviving channel's $\mu_z$ is interpreted as the extracted parameter and is fed, together with the state time series, into the reservoir computer's parameter input. The prediction protocol is to apply a small change $\Delta z$ beyond the training range and let the reservoir evolve autonomously; if the generated attractor is no longer the healthy one, a critical transition is flagged.

What would settle it

Drive the trained reservoir computer with the VAE-extracted latent parameter at a value well inside the collapse regime for the Lorenz system (\rho beyond the reported \rho_c \approx 24). If, over an ensemble of reservoir realizations, the fraction of runs that produce the pre-crisis chaotic attractor is comparable to the fraction at healthy \rho, then the extrapolation from training-range latent values to unseen values is not working, and the central claim would fail. A second decisive check is to run the same experiment with the true parameter \rho injected instead of the VAE's z: if the true-parameter reservoir predicts the crisis but the VAE-driven one does not, the error is in the parameter extraction rather than in the reservoir's anticipation capability.

Watch

Extended reading notes

Core claim

The central claim is that the bifurcation parameter and its variations can be faithfully extracted from time-series data with no prior knowledge of the parameter, and that a reservoir computer driven by the extracted latent parameter can then anticipate the critical transition. Concretely, the VAE's encoder outputs a mean and variance for each latent channel; a channel that shows large variance of the mean across inputs and small mean variance within inputs is identified as the true hidden parameter, and its mean is used as the effective bifurcation parameter. In the Lorenz examples the extracted latent variable is linearly related to the true parameter, and the reservoir's predicted critical point concentrates near the true boundary crisis value ($\rho_c \approx 24.06$, predicted $\approx 24 \pm 0.5$). For the Kuramoto-Sivashinsky system, about 98% of reservoir realizations predict the critical point within 10% relative error. The paper also shows that the same framework handles two independent bifurcation parameters, partial state observation via time-delay embedding in the decoder, and an ecosystem model whose equations are not sparse.

Load-bearing premise

The load-bearing premise is that a reservoir computer trained only with latent parameters from the normal regime can extrapolate to latent values it has never seen and still correctly label the true system's state as healthy or collapsed.

Editorial extensions

If this is right

Critical transitions can be anticipated from time series alone, without measuring or knowing the bifurcation parameter, removing a key obstacle to practical early-warning systems.
The VAE's latent variable can itself serve as an effective bifurcation parameter, so monitoring and prediction can proceed in latent space without mapping back to physical units.
The same framework extends to multiple independent bifurcation parameters and to partial state observation, broadening it beyond single-parameter, full-state idealizations.
Spatiotemporal systems with sharp collapse thresholds, such as the Kuramoto-Sivashinsky equation, are within the method's reach, with 98% of reservoir realizations predicting the critical point within 10% relative error.
Because the approach does not rely on sparse equation structure, it applies to systems like a chaotic food chain where equation-discovery methods would fail.

Reading between the lines

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

Beyond the paper, the channel-selection statistic could be monitored in time as an online early-warning indicator: a latent channel becoming informative might precede the statistical signatures usually used for tipping-point detection.
Beyond the paper, the linear relationship between the latent variable and the true parameter, found in every test system, suggests the VAE is learning a coordinate reparametrization of the parameter manifold; if that is generic, one could map out entire bifurcation diagrams by sweeping the latent variable.
Beyond the paper, the central risk is that the reservoir is trained and tested on the same dynamical family; a stress test on a different system topology with the same VAE would clarify how much of the success is due to the VAE's extraction versus the reservoir's interpolation within a familiar dynamical climate.

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.

Desk Editor's Note

read the letter

Quick take: the paper does something genuinely new—it replaces the known bifurcation parameter in parameter-driven reservoir computing with a VAE-extracted latent variable, and shows on four simulated systems that the resulting pipeline can anticipate a boundary crisis or spatiotemporal collapse. The demonstrations are clean and the system coverage is decent (Lorenz with one and two parameters, Kuramoto-Sivashinsky, partial observation, a food-chain model). The latent-channel heuristic—look for high variance in µ_z and low mean σ²_z—works in all cases, and the predicted critical points land near the true values. That is real evidence.

The soft spot is exactly where the stress-test note puts it: the load-bearing extrapolation is untested. The reservoir is trained only on time series from the normal regime, then driven with z values outside that regime; whether it correctly classifies collapse depends on both the VAE's z-p extrapolation and the reservoir's own generalization. The paper reports no baseline where the true parameter is fed to the reservoir, and no control using a collapsed latent channel. The affine maps (C1, C2, C) that convert z to physical units are fit to ground-truth parameters, so the physical-unit predictions are calibrated, not predicted. The paper's own Appendix A.2 warns that black-box predictors don't extrapolate accurately; nothing here shows this pipeline is the exception. That said, the paper is honest about the calibration role: it says mapping back to physical units is only for validation, and the latent variable is the effective parameter in practice. So the flaw is not fatal, but the claim of 'faithfully extracted' is stronger than what the evidence supports.

Other soft spots are milder: no code or data, a few free hyperparameters tuned per system, and the two-parameter case shows FPR ≈ 23%, which is worth a closer look. The citation pattern is fine; the relevant prior work (Lu et al. 2020, Kong et al. 2021) is cited.

Who this is for: nonlinear-dynamics folks working on ML-based anticipation. It deserves a serious referee—the core idea is worth testing properly with the right baselines. I'd want the authors to add a true-parameter baseline, a collapse-channel control, and an out-of-training z-p fit before publication.

Referee Report

4 major / 6 minor

Summary. This manuscript proposes an end-to-end machine-learning pipeline for anticipating critical transitions without explicit knowledge of the bifurcation parameter. A variational autoencoder with a convolutional encoder and a feedback decoder is trained on time series from the healthy regime; the active latent channel(s) are interpreted as an inferred effective bifurcation parameter z. A parameter-driven reservoir computer is then trained on pairs of time series and inferred z values, and is tested by extrapolating z beyond the training range to decide whether the predicted long-term dynamics are healthy or collapsed. The method is demonstrated on the Lorenz system with one and two parameters, the Kuramoto-Sivashinsky system, a partial-observation Lorenz case, and a chaotic food-chain model. The paper claims that the framework relaxes the requirement of knowing the bifurcation parameter, since the parameter and its variations can be faithfully extracted from data.

Significance. If the central claim holds, the contribution is significant: it would remove a key practical requirement of parameter-driven reservoir computing and extend unsupervised parameter extraction to the anticipation of critical transitions, including spatiotemporal systems. The paper has clear strengths: it treats several non-trivial testbeds, reports ensemble statistics over many reservoir realizations, gives a full hyperparameter table, and explicitly discusses the limitations of sparse-optimization and black-box extrapolation in Appendix A. The partial-observation and multi-parameter extensions are useful demonstrations. However, the evidence currently does not fully separate the VAE's extraction quality from the reservoir's extrapolation behavior, and the validation relies on affine maps fitted to ground-truth parameter values. Those gaps are load-bearing for the advertised 'no prior knowledge' claim.

major comments (4)

[Abstract; main text, Fig. 3(c)-(d); Appendix D] The central claim that the framework requires no prior knowledge of the bifurcation parameter is not supported end-to-end, because the reported predictions are converted into physical-parameter units using affine constants C1 and C2 (Fig. 3(d)) and the matrix C (Appendix D) that are fit to ground-truth parameter values from the training data. The authors say this mapping is 'solely for validating the proposed method,' but this means the validation itself is not free of ground-truth parameter information. In a real deployment, no ground truth p is available to fit C1 and C2; the paper does not show that a latent-space threshold or other calibration-free decision rule yields the same critical-point accuracy. Please either provide a validation conducted entirely in latent units, or show that the calibration can be obtained from a known safe operating interval only.
[Appendix C; Figs. 3(d), 4(e), 7(c), 8(d)] The testing protocol applies a parameter change Δz outside the training range and judges collapse by the reservoir's autonomous output. Because the reservoir also receives the raw time series as input, the reported accuracy could partly arise from the reservoir recognizing unhealthy temporal statistics directly, rather than from the VAE's inferred z. There is no baseline in which the reservoir is driven by the true parameter p, and no ablation with a constant, random, or corrupted latent channel. Such baselines are needed to separate VAE extraction error from reservoir extrapolation error and to substantiate the claim that the latent parameter is the information carrier.
[Appendix C; Appendix A.2; Figs. 3(a), 4, 8(a)] The success of the method hinges on extrapolation: the reservoir is trained only on the healthy regime but must classify states for z values beyond the training interval. The paper provides no continuity or stability argument for the latent dynamics in this regime, and it does not report how far the predicted critical latent value lies outside the training range or how the predicted critical point varies with the extrapolation distance. This is especially important because Appendix A.2 itself warns that a black-box predictor does not provide accurate extrapolation of dynamical changes; the paper should explain why the VAE-reservoir pipeline is not subject to that warning, and support the explanation with additional experiments such as a systematic sweep of the testing range with and without true-parameter driving.
[Figs. 3(c), 4(d), 7(b), 8(c); Appendix D, Fig. 6(c)] The claim that the parameter is 'faithfully extracted' is based on linear fits of z versus p, but no goodness-of-fit statistics, scatter, or error bars are reported for these relations. In the two-parameter Lorenz case the extracted latent channels are linear combinations of the two true parameters (Fig. 6(c)), so it is not clear whether the VAE recovers the physical parameters or merely an invertible mixture; the FPR and FNR reported in Appendix D are also conditioned on the ground-truth-fitted matrix C. Please report R² and residual statistics for the one-parameter fits, and analyze the conditioning and sensitivity of C for the two-parameter case.

minor comments (6)

[Appendix F, Eq. (F1)] The consumer equation contains a typo, '− −xpypP C', which should be a single minus sign.
[References [9], main text] The text repeatedly refers to Supplementary sections S1 through S7, but the arXiv posting does not include the supplementary file; readers cannot check the details of the methods or hyperparameters. Please ensure the supplementary material is posted or remove the pointers.
[Figs. 3(c) and 3(d)] The notation ρ* is used both for the extracted latent parameter mapped into physical units (Fig. 3(c)) and for the predicted critical point (Fig. 3(d)); please use distinct symbols to avoid confusion.
[Appendix G] The statement that 'the number of hidden layers is fixed at 32' is likely a typo; a VAE with 32 hidden layers would be unusual, and the surrounding discussion of batch size and learning rate suggests hidden units were meant. Please clarify.
[Fig. 6(d) caption] The caption says 'green red dots', which should read 'green and red dots'.
[Appendix B, Eq. (B1)] The objective is described as a 'minimax optimization problem,' but the loss in Eq. (B1) is minimized; please rephrase.

Circularity Check

1 steps flagged · score 3.0 of 10

Physical-unit critical-point predictions are calibrated by affine maps fitted to ground-truth parameters, but the latent-space transition detection itself is unsupervised and independent.

fitted input called prediction [Main text, Lorenz single-parameter results, Fig. 3(c)-(d); repeated in Fig. 4(e), Fig. 7(c), and Appendix D.]
"Figure 3(d) shows a histogram of the predicted critical point ρ∗, where the linear parameter transformation ρ∗ = C1 ˆz + C2 is used with C1 = 4.3 and C2 = 28.5 so as to map the z values to the real domain ρ∗. The resulting distribution centers about the ground truth ρc ≈ 24.06."

The constants C1 and C2 are the coefficients of the linear fit shown in Fig. 3(c), where the VAE-extracted z is regressed against the known ground-truth parameter ρ on the training interval. The reported physical-unit critical point is therefore obtained by applying a supervised calibration to the latent-space reservoir threshold. The agreement with ρc is partly inherited from this calibration rather than being a purely unsupervised prediction of the physical parameter value. The latent-space transition detection itself is not forced by this construction, so the circularity is partial: the calibration step is not equivalent to the central claim, but the headline numeric predictions in physical units are not fully self-contained.

full rationale

The VAE latent parameter is learned from reconstruction loss without ground-truth parameter supervision, and the reservoir is driven by extrapolated latent values, so the core unsupervised anticipation claim has independent content. The only partially circular element is the conversion of the latent threshold into a physical-unit critical point through affine maps fitted to the true parameters. Because the paper explicitly states that this mapping is solely for validation, this is a calibration issue rather than a derivation-level circularity. Self-citations to the authors' earlier parameter-driven reservoir computing work are not load-bearing here, since that prior method is used as a component and its validity is not the paper's central claim.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a learned latent variable, a supervised calibration to physical units, and the assumed extrapolation capability of reservoir computing. These supports are internal to the paper; no independent data, code, or formal analysis is provided.

free parameters (6)

Affine calibration constants C1, C2 (Lorenz, one parameter) = C1 = 4.3, C2 = 28.5
Fit to map the VAE latent variable z to the physical parameter rho for reporting the predicted critical point; uses ground-truth parameter labels.
Affine calibration constants C1, C2 (Kuramoto-Sivashinsky) = C1 = -4.5, C2 = 191.13
Fit to map z to phi for reporting the predicted critical point; uses ground-truth parameter labels.
Affine calibration constants C1, C2 (Lorenz, partial observation) = C1 = -3.5, C2 = 29.3
Fit to map z to rho in the partial-observation example; uses ground-truth parameter labels.
Affine calibration matrix C (Lorenz, two parameters) = [[1.74, -2.45, 30.15], [-0.48, -0.33, 1.72]]
Fit to map the two latent variables to rho and beta for reporting predicted behavior; uses ground-truth parameter pairs.
Number of VAE latent channels = 5 for all examples
Chosen by hand; must be larger than the number of true bifurcation parameters for the channel-selection heuristic to work.
Bayesian-optimized hyperparameters (VAE and reservoir) = Table I, Sets A-E (e.g., learning rate 1e-3 to 5e-3, spectral radius 0.55 to 2.3, reservoir size 800 to 4000)
Selected per example against validation error; the claimed performance depends on these choices, and no code is provided to reproduce them.

assumptions (5)

domain assumption Reservoir computer trained on normal-regime data extrapolates correctly to parameter values outside the training range.
Required for forecasting collapse at z beyond the training interval; tested empirically but not proven. Location: Appendix C and the testing protocols in the main text.
domain assumption The VAE latent variable z is identifiable and affinely related to the true bifurcation parameter.
The paper assumes the active latent channel corresponds one-to-one to the physical parameter; supported by examples, not by an identifiability guarantee. Location: Fig. 2 and Figs. 3(b)-(c).
domain assumption Time-series data are available from several distinct parameter values in the normal regime.
Training requires sampling the bifurcation parameter over an interval, e.g., 'randomly sample rho in [25,35]'; real-world data may not satisfy this. Location: training description in the main text.
standard math Takens' delay-coordinate embedding is valid for the partial-observation scenario.
Used to justify feeding delayed coordinates to the decoder under partial observation. Location: Appendix E.
standard math Reservoir computing hyperparameters satisfy the echo state property for the chosen systems.
The reservoir is assumed to be in a stable regime so its predictions are meaningful; this is standard practice but not verified in the paper. Location: Appendix C.

invented entities (1)

VAE latent variable z as an effective bifurcation parameter
purpose: Serves as the parameter input to the reservoir computer and as a monitoring coordinate for anticipating critical transitions when the physical parameter is unknown.
z is a learned latent quantity; the paper establishes only an empirical correlation with the true parameter in training data, not an identifiability guarantee or an external falsifiable prediction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Unsupervised learning for anticipating critical transitions." pith.science (2026). https://pith.science/paper/SZEEZKB4

@misc{pith2026250101579,
  author       = {Pith},
  title        = {Pith review of: Unsupervised learning for anticipating critical transitions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SZEEZKB4}},
  note         = {Machine review of arXiv:2501.01579}
}

read the original abstract

For anticipating critical transitions in complex dynamical systems, the recent approach of parameter-driven reservoir computing requires explicit knowledge of the bifurcation parameter. We articulate a framework combining a variational autoencoder (VAE) and reservoir computing to address this challenge. In particular, the driving factor is detected from time series using the VAE in an unsupervised-learning fashion and the extracted information is then used as the parameter input to the reservoir computer for anticipating the critical transition. We demonstrate the power of the unsupervised learning scheme using prototypical dynamical systems including the spatiotemporal Kuramoto-Sivashinsky system. The scheme can also be extended to scenarios where the target system is driven by several independent parameters or with partial state observations.

Figures

Figures reproduced from arXiv: 2501.01579 by the authors.

**Figure 2.** FIG. 2. Principle of VAE identification of the bifurcation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** FIG. 4. Predicting critical transition in the nonlinear wave [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: FIG. 5. VAEs and parameter-driven reservoir computing. (a) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: FIG. 6. Unsupervised learning for anticipating critical transition of Lorenz system with two independent bifurcation parameters [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: FIG. 7. Unsupervised-learning based prediction critical tran [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: FIG. 8. Anticipating a critical transition in the chaotic food-chain system (F1). (a) Bifurcation diagram with [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Tailored Forecasting from Short Time Series via Meta-learning
cs.LG 2025-01 conditional novelty 6.0 of 10

METAFORS maps short unlabeled time series to tailored forecaster parameters and cold-start states, enabling accurate short-term and climate forecasts for unseen chaotic systems from as few as two data points.

Reference graph

Works this paper leans on

83 extracted references · 67 canonical work pages · cited by 1 Pith paper

[1]

Sparse optimization for finding the governing equations from data An earlier approach to anticipating critical transitions was articulated [10] in 2011, which is based on finding the governing equations of the system from data [34– 36]. The idea is that the equations of certain nonlin- ear dynamical systems possess a sparse structure: they contain a small...

2011
[2]

normal” in the sense that its dynamical behavior ex- hibits “healthy

Limitation of previous parameter-driven reservoir computing A limitation of parameter-driven reservoir computing is that it requires the knowledge of the bifurcation or con- trol parameter that may change with time. In particular, during the standard training process [13], the bifurcation parameter values need to be injected into the neural net- work toge...
[3]

Hastings, K

A. Hastings, K. C. Abbott, K. Cuddington, T. Fran- cis, G. Gellner, Y.-C. Lai, A. Morozov, S. Petrovskii, K. Scranton, and M. L. Zeeman, Transient phenomena in ecology, Science 361, eaat6412 (2018)

2018
[4]

T. M. Lenton, J. Rockstr¨ om, O. Gaffney, S. Rahmstorf, K. Richardson, W. Steffen, and H. J. Schellnhuber, Cli- mate tipping points—too risky to bet against, Nature 575, 592 (2019)

work page 2019
[5]

Dobson and H.-D

I. Dobson and H.-D. Chiang, Towards a theory of voltage collapse in electric power systems, Sys. Cont. Lett. 13, 12 253 (1989)

work page 1989
[6]

Scheffer, Critical Transitions in Nature and Society, Princeton Studies in Complexity (Princeton University Press, 2020)

M. Scheffer, Critical Transitions in Nature and Society, Princeton Studies in Complexity (Princeton University Press, 2020)

work page 2020
[7]

S. M. O’Regan, E. B. O’Dea, P. Rohani, and J. M. Drake, Transient indicators of tipping points in infectious dis- eases, J. R. Soc. Interface. 17, 20200094 (2020)

work page 2020
[8]

E. H. van Nes, B. M. Arani, A. Staal, B. van der Bolt, B. M. Flores, S. Bathiany, and M. Scheffer, What do you mean, ‘tipping point’ ?, Trends Ecol Evol. 31, 902–904 (2016)

work page 2016

Show all 83 references

[10]

Scheffer, S

M. Scheffer, S. R. Carpenter, T. M. Lenton, J. Bas- compte, W. Brock, V. Dakos, J. van de Koppel, I. A. van de Leemput, S. A. Levin, E. H. van Nes, M. Pascual, and J. Vandermeer, Anticipating critical transitions, Sci- ence 338, 344–348 (2012)

2012
[11]

It is helpful but not essen- tial for understanding the main results of the paper

Supplementary Information provides details elaborating the results in the main text. It is helpful but not essen- tial for understanding the main results of the paper. It contains the following: (1) background materials on pre- vious approaches to anticipating critical transit...
[12]

W.-X. Wang, R. Yang, Y.-C. Lai, V. Kovanis, and C. Grebogi, Predicting catastrophes in nonlinear dynami- cal systems by compressive sensing, Phys. Rev. Lett.106, 154101 (2011)

2011
[13]

S. H. Lim, L. Theo Giorgini, W. Moon, and J. S. Wet- tlaufer, Predicting critical transitions in multiscale dy- namical systems using reservoir computing, Chaos 30 (2020)

2020
[14]

T. M. Bury, R. Sujith, I. Pavithran, M. Scheffer, T. M. Lenton, M. Anand, and C. T. Bauch, Deep learning for early warning signals of tipping points, Proc. Natl. Acad. Sci. (USA) 118, e2106140118 (2021)

2021
[15]

Kong, H.-W

L.-W. Kong, H.-W. Fan, C. Grebogi, and Y.-C. Lai, Ma- chine learning prediction of critical transition and system collapse, Phys. Rev. Res. 3, 013090 (2021)

2021
[16]

J. Z. Kim, Z. Lu, E. Nozari, G. J. Pappas, and D. S. Bas- sett, Teaching recurrent neural networks to infer global temporal structure from local examples, Nat. Mach. In- tell. 3, 316 (2021)

2021
[17]

Fan, L.-W

H. Fan, L.-W. Kong, Y.-C. Lai, and X. Wang, Anticipat- ing synchronization with machine learning, Phys. Rev. Res. 3, 023237 (2021)

2021
[18]

L.-W. Kong, H. Fan, C. Grebogi, and Y.-C. Lai, Emer- gence of transient chaos and intermittency in machine learning, J. Phys. Complex. 2, 035014 (2021)

2021
[19]

L.-W. Kong, Y. Weng, B. Glaz, M. Haile, and Y.-C. Lai, Reservoir computing as digital twins for nonlinear dy- namical systems, Chaos 33, 033111 (2023)

2023
[20]

echo state

H. Jaeger, The “echo state” approach to analysing and training recurrent neural networks-with an erratum note, GMD 148, 13 (2001)

2001
[21]

Maass, T

W. Maass, T. Natschl¨ ager, and H. Markram, Real-time computing without stable states: A new framework for neural computation based on perturbations, Neural Com- put. 14, 2531 (2002)

2002
[22]

Pathak, B

J. Pathak, B. Hunt, M. Girvan, Z. Lu, and E. Ott, Model- free prediction of large spatiotemporally chaotic systems from data: A reservoir computing approach, Phys. Rev. Lett. 120, 024102 (2018)

2018
[23]

E. Bollt, On explaining the surprising success of reservoir computing forecaster of chaos? the universal machine learning dynamical system with contrast to var and dmd, Chaos 31, 013108 (2021)

2021
[24]

D. J. Gauthier, E. Bollt, A. Griffith, and W. A. Barbosa, Next generation reservoir computing, Nat. Commun. 12, 5564 (2021)

2021
[25]

Zhai, L.-W

Z.-M. Zhai, L.-W. Kong, and Y.-C. Lai, Emergence of a resonance in machine learning, Phys. Rev. Res. 5, 033127 (2023)

2023
[26]

L.-W. Kong, G. A. Brewer, and Y.-C. Lai, Reservoir- computing based associative memory and itinerancy for complex dynamical attractors, Nat. Commun. 15, 2815 (2024)

2024
[27]

M. Yan, C. Huang, P. Bienstman, P. Tino, W. Lin, and J. Sun, Emerging opportunities and challenges for the future of reservoir computing, Nat. Commun. 15, 2056 (2024)

2024
[28]

Xiao, L.-W

R. Xiao, L.-W. Kong, Z.-K. Sun, and Y.-C. Lai, Predict- ing amplitude death with machine learning, Phys. Rev. E 104, 014205 (2021)

2021
[29]

Patel, D

D. Patel, D. Canaday, M. Girvan, A. Pomerance, and E. Ott, Using machine learning to predict statistical properties of non-stationary dynamical processes: Sys- tem climate, regime transitions, and the effect of stochas- ticity, Chaos 31, 033149 (2021)

2021
[30]

Panahi, L.-W

S. Panahi, L.-W. Kong, M. Moradi, Z.-M. Zhai, B. Glaz, M. Haile, and Y.-C. Lai, Machine-learning prediction of tipping, arXiv preprint arXiv:2402.14877 (2024)

2024 arXiv
[31]

P. Y. Lu, S. Kim, and M. Soljaˇ ci´ c, Extracting inter- pretable physical parameters from spatiotemporal sys- tems using unsupervised learning, Phys. Rev. X 10, 031056 (2020)

2020
[32]

E. N. Lorenz, Deterministic nonperiodic flow, J. Atmos. Sci. 20, 130 (1963)

1963
[33]

Kuramoto, Diffusion-induced chaos in reaction sys- tems, Prog

Y. Kuramoto, Diffusion-induced chaos in reaction sys- tems, Prog. Theo. Phys. Supp. 64, 346 (1978)

1978
[34]

G. I. Sivashinsky, On flame propagation under conditions of stoichiometry, SIAM J. Appl. Math. 39, 67 (1980)

1980
[35]

Grebogi, E

C. Grebogi, E. Ott, and J. A. Yorke, Crises, sudden changes in chaotic attractors, and transient chaos, Phys- ica D 7, 181–200 (1983)

1983
[36]

J. P. Crutchfield and B. McNamara, Equations of motion from a data series, Complex Sys. 1, 417 (1987)

1987
[37]

E. M. Bollt, Controlling chaos and the inverse frobenius- perron problem: global stabilization of arbitrary invari- ant measures, Int. J. Bif. Chaos 10, 1033 (2000)

2000
[38]

Yao and E

C. Yao and E. M. Bollt, Modeling and nonlinear parame- ter estimation with Kronecker product representation for coupled oscillators and spatiotemporal systems, Physica D 227, 78 (2007)

2007
[39]

O. E. R¨ ossler, An equation for continuous chaos, Phys. Lett. A 57, 397 (1976)

1976
[40]

H´ enon, A two-dimensional mapping with a strange attractor, Commun

M. H´ enon, A two-dimensional mapping with a strange attractor, Commun. Math. Phys. 50, 69 (1976). 13

1976
[41]

Cand` es, J

E. Cand` es, J. Romberg, and T. Tao, Robust uncertainty principles: exact signal reconstruction from highly in- complete frequency information, IEEE Trans. Info. The- ory 52, 489 (2006)

2006
[42]

Cand` es, J

E. Cand` es, J. Romberg, and T. Tao, Stable signal re- covery from incomplete and inaccurate measurements, Comm. Pure Appl. Math. 59, 1207 (2006)

2006
[43]

Donoho, Compressed sensing, IEEE Trans

D. Donoho, Compressed sensing, IEEE Trans. Info. The- ory 52, 1289 (2006)

2006
[44]

R. G. Baraniuk, Compressed sensing, IEEE Signal Pro- cess. Mag. 24, 118 (2007)

2007
[45]

Cande` s and M

E. Cande` s and M. Wakin, An introduction to compressive sampling, IEEE Signal Process. Mag. 25, 21 (2008)

2008
[46]

Wang, Y.-C

W. Wang, Y.-C. Lai, and C. Grebogi, Data based identi- fication and prediction of nonlinear and complex dynam- ical systems, Phys. Rep. 644, 1 (2016)

2016
[47]

Lai, Finding nonlinear system equations and com- plex network structures from data: A sparse optimization approach, Chaos 31, 082101 (2021)

Y.-C. Lai, Finding nonlinear system equations and com- plex network structures from data: A sparse optimization approach, Chaos 31, 082101 (2021)

2021
[48]

W.-X. Wang, R. Yang, Y.-C. Lai, V. Kovanis, and M. A. F. Harrison, Time-series-based prediction of com- plex oscillator networks via compressive sensing, EPL (Europhys. Lett.) 94, 48006 (2011)

2011
[49]

Wang, Y.-C

W.-X. Wang, Y.-C. Lai, C. Grebogi, and J.-P. Ye, Net- work reconstruction based on evolutionary-game data via compressive sensing, Phys. Rev. X 1, 021021 (2011)

2011
[50]

R.-Q. Su, X. Ni, W.-X. Wang, and Y.-C. Lai, Forecasting synchronizability of complex networks from data, Phys. Rev. E 85, 056220 (2012)

2012
[51]

Su, W.-X

R.-Q. Su, W.-X. Wang, and Y.-C. Lai, Detecting hidden nodes in complex networks from time series, Phys. Rev. E 85, 065201 (2012)

2012
[52]

Su, Y.-C

R.-Q. Su, Y.-C. Lai, X. Wang, and Y.-H. Do, Uncover- ing hidden nodes in complex networks in the presence of noise, Sci. Rep. 4, 3944 (2014)

2014
[53]

Shen, W.-X

Z. Shen, W.-X. Wang, Y. Fan, Z. Di, and Y.-C. Lai, Re- constructing propagation networks with natural diversity and identifying hidden sources, Nat. Commun. 5, 4323 (2014)

2014
[54]

Scheffer, Ecology of Shallow Lakes(Springer Science & Business Media, 2004)

M. Scheffer, Ecology of Shallow Lakes(Springer Science & Business Media, 2004)

2004
[55]

Scheffer, Complex systems: foreseeing tipping points, Nature 467, 411 (2010)

M. Scheffer, Complex systems: foreseeing tipping points, Nature 467, 411 (2010)

2010
[56]

D. B. Wysham and A. Hastings, Regime shifts in ecolog- ical systems can occur with no warning, Ecol. Lett. 13, 464 (2010)

2010
[57]

J. M. Tylianakis and C. Coux, Tipping points in ecolog- ical networks, Trends. Plant. Sci. 19, 281 (2014)

2014
[58]

Jiang, Z.-G

J. Jiang, Z.-G. Huang, T. P. Seager, W. Lin, C. Grebogi, A. Hastings, and Y.-C. Lai, Predicting tipping points in mutualistic networks through dimension reduction, Proc. Nat. Acad. Sci. (USA) 115, E639 (2018)

2018
[59]

B. Yang, M. Li, W. Tang, W. Liu, S. Zhang, L. Chen, and J. Xia, Dynamic network biomarker indicates pul- monary metastasis at the tipping point of hepatocellular carcinoma, Nat. Commun. 9, 678 (2018)

2018
[60]

Jiang, A

J. Jiang, A. Hastings, and Y.-C. Lai, Harnessing tipping points in complex ecological networks, J. R. Soc. Interface 16, 20190345 (2019)

2019
[61]

Meng, Y.-C

Y. Meng, Y.-C. Lai, and C. Grebogi, Tipping point and noise-induced transients in ecological networks, J. R. Soc. Interface. 17, 20200645 (2020)

2020
[62]

Meng, Y.-C

Y. Meng, Y.-C. Lai, and C. Grebogi, The fundamental benefits of multiplexity in ecological networks, J. R. Soc. Interface 19, 20220438 (2022)

2022
[63]

Scheffer, J

M. Scheffer, J. Bascompte, W. A. Brock, V. Brovkin, S. R. Carpenter, V. Dakos, H. Held, E. H. Van Nes, M. Rietkerk, and G. Sugihara, Early-warning signals for critical transitions, Nature 461, 53 (2009)

2009
[64]

J. M. Drake and B. D. Griffen, Early warning signals of extinction in deteriorating environments, Nature 467, 456 (2010)

2010
[65]

Boettiger and A

C. Boettiger and A. Hastings, Quantifying limits to de- tection of early warning for critical transitions, J. R. Soc. Interface 9, 2527 (2012)

2012
[66]

L. Chen, R. Liu, Z.-P. Liu, M. Li, and K. Aihara, De- tecting early-warning signals for sudden deterioration of complex diseases by dynamical network biomarkers, Sci. Rep. 2, 342 (2012)

2012
[67]

L. Dai, D. Vorselen, K. S. Korolev, and J. Gore, Generic indicators for loss of resilience before a tipping point lead- ing to population collapse, Science 336, 1175 (2012)

2012
[68]

Boettiger, N

C. Boettiger, N. Ross, and A. Hastings, Early warning signals: the charted and uncharted territories, Theor. Ecol. 6, 255 (2013)

2013
[69]

I. A. van de Leemput, M. Wichers, A. O. Cramer, D. Borsboom, F. Tuerlinckx, P. Kuppens, E. H. van Nes, W. Viechtbauer, E. J. Giltay, S. H. Aggen, et al., Critical slowing down as early warning for the onset and termi- nation of depression, Proc. Natl. Acad. Sci. (USA) 111, 87 (2014)

2014
[70]

Boers, Early-warning signals for dansgaard-oeschger events in a high-resolution ice core record, Nat

N. Boers, Early-warning signals for dansgaard-oeschger events in a high-resolution ice core record, Nat. Commun. 9, 1 (2018)

2018
[71]

Zhang, P

Y. Zhang, P. Tino, A. Leonardis, and K. Tang, A survey on neural network interpretability, IEEE Trans. Emerg. Top. Comput. Intell. 5, 726–742 (2021)

2021
[72]

Rudin, Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead, Nat

C. Rudin, Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead, Nat. Mach. Intell. 1, 206–215 (2019)

2019
[73]

O. Alao, P. Y. Lu, and M. Soljacic, Discovering dynam- ical parameters by interpreting echo state networks, in NeurIPS 2021 AI for Science Workshop(2021)

2021
[74]

Goodfellow, Y

I. Goodfellow, Y. Bengio, and A. Courville, Deep Learn- ing, Adaptive Computation and Machine Learning series (MIT Press, 2016)

2016
[75]

D. P. Kingma and M. Welling, Auto-encoding variational bayes, arXiv preprint arXiv:1312.6114 (2013)

2013 arXiv
[76]

D. J. Rezende, S. Mohamed, and D. Wierstra, Stochastic backpropagation and approximate inference in deep gen- erative models, in International Conference on Machine Learning (PMLR, 2014) pp. 1278–1286

2014
[77]

Manjunath and H

G. Manjunath and H. Jaeger, Echo state property linked to an input: Exploring a fundamental characteristic of recurrent neural networks, Neur. Comp. 25, 671 (2013)

2013
[78]

F. Takens, Detecting strange attractors in turbulence, in Dynamical Systems and Turbulence, Warwick 1980: pro- ceedings of a symposium held at the University of War- wick 1979/80 (Springer, 2006) pp. 366–381

1980
[79]

McCann and P

K. McCann and P. Yodzis, Nonlinear dynamics and pop- ulation disappearances, Am. Nat. 144, 873–879 (1994)

1994
[80]

Bergstra, R

J. Bergstra, R. Bardenet, Y. Bengio, and B. K´ egl, Algo- rithms for hyper-parameter optimization, in Advances in Neural Information Processing Systems, Vol. 24, edited by J. Shawe-Taylor, R. Zemel, P. Bartlett, F. Pereira, and K. Weinberger (Curran Associates, Inc., 2011)

2011
[81]

Bergstra and Y

J. Bergstra and Y. Bengio, Random search for hyper- parameter optimization., J. Mach. Learn. Res. 13 (2012). 14

2012
[82]

Zoph and Q

B. Zoph and Q. V. Le, Neural architecture search with reinforcement learning, arXiv preprint arXiv:1611.01578 (2016)

2016 arXiv
[83]

C. K. Williams and C. E. Rasmussen, Gaussian Pro- cesses for Machine Learning, Vol. 2 (MIT press Cam- bridge, MA, 2006)

2006
[84]

Breiman, Random forests, Mach

L. Breiman, Random forests, Mach. Learn. 45, 5 (2001)

2001

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Sparse optimization for finding the governing equations from data An earlier approach to anticipating critical transitions was articulated [10] in 2011, which is based on finding the governing equations of the system from data [34– 36]. The idea is that the equations of certain nonlin- ear dynamical systems possess a sparse structure: they contain a small...

2011

[2] [2]

normal” in the sense that its dynamical behavior ex- hibits “healthy

Limitation of previous parameter-driven reservoir computing A limitation of parameter-driven reservoir computing is that it requires the knowledge of the bifurcation or con- trol parameter that may change with time. In particular, during the standard training process [13], the bifurcation parameter values need to be injected into the neural net- work toge...

[3] [3]

Hastings, K

A. Hastings, K. C. Abbott, K. Cuddington, T. Fran- cis, G. Gellner, Y.-C. Lai, A. Morozov, S. Petrovskii, K. Scranton, and M. L. Zeeman, Transient phenomena in ecology, Science 361, eaat6412 (2018)

2018

[4] [4]

T. M. Lenton, J. Rockstr¨ om, O. Gaffney, S. Rahmstorf, K. Richardson, W. Steffen, and H. J. Schellnhuber, Cli- mate tipping points—too risky to bet against, Nature 575, 592 (2019)

work page 2019

[5] [5]

Dobson and H.-D

I. Dobson and H.-D. Chiang, Towards a theory of voltage collapse in electric power systems, Sys. Cont. Lett. 13, 12 253 (1989)

work page 1989

[6] [6]

Scheffer, Critical Transitions in Nature and Society, Princeton Studies in Complexity (Princeton University Press, 2020)

M. Scheffer, Critical Transitions in Nature and Society, Princeton Studies in Complexity (Princeton University Press, 2020)

work page 2020

[7] [7]

S. M. O’Regan, E. B. O’Dea, P. Rohani, and J. M. Drake, Transient indicators of tipping points in infectious dis- eases, J. R. Soc. Interface. 17, 20200094 (2020)

work page 2020

[8] [8]

E. H. van Nes, B. M. Arani, A. Staal, B. van der Bolt, B. M. Flores, S. Bathiany, and M. Scheffer, What do you mean, ‘tipping point’ ?, Trends Ecol Evol. 31, 902–904 (2016)

work page 2016