Echoes Before Collapse: Deep Learning Detection of Flickering in Complex Systems

Yazdan Babazadeh Maghsoodlo, Madhur Anand, Chris T. Bauch · 2025

Read Echoes Before Collapse: Deep Learning Detection of Flickering in Complex Systems by Yazdan Babazadeh Maghsoodlo, Madhur Anand, Chris T. Bauch (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-04683.txt.

Introduction

Flickering, the noisy jumping between two stable states, can warn that a system is losing resilience. Yazdan Babazadeh Maghsoodlo, lead author with two co-authors, show deep learning can detect it in a 2025 paper. Their models combine convolutional and long short-term memory networks, trained on synthetic time series with added noise.

Although training used simplified dynamics, the models generalise to varied stochastic systems. They detect flickering in real records, including dormouse body temperature and palaeoclimate proxies from the African Humid Period. The authors offer the approach as a flexible early warning tool for climate, ecological, and other complex systems.

Catalog record

CREATORYazdan Babazadeh Maghsoodlo, Madhur Anand, Chris T. Bauch
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-04683

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Echoes Before Collapse: Deep Learning Detection of Flickering in Complex Systems Yazdan Babazadeh Maghsoodlo, Madhur Anand, Chris T. Bauch Abstract 1 Abstract Deep learning offers powerful tools for anticipating tipping points in complex systems, yet its potential for detecting flickering (noise-driven switching between coexisting stable states) remains unexplored. Flickering is a hallmark of reduced resilience in climate systems, ecosystems, financial markets, and other systems. It can precede critical regime shifts that are highly impactful but difficult to predict. Here we show that convolutional–long short-term memory (CNN–LSTM) models, trained on synthetic time series generated from simple polynomial functions with additive noise, can accurately identify flickering patterns. Despite being trained on simplified dynamics, our models generalize to diverse stochastic systems and reliably detect flickering in empirical datasets, including dormouse body temperature records and palaeoclimate proxies from the African Humid Period. These findings demonstrate that deep learning can extract early warning signals from noisy, nonlinear time series, providing a flexible framework for identifying instability across a wide range of dynamical systems. 2 Introduction Sudden and irreversible shifts, tipping points [ 44 , 23 , 26 , 39 ] , can arise in systems as varied as climate [ 8 , 38 ] , ecosystems [ 24 , 31 , 41 ] , financial markets [ 7 , 33 ] , and public health [ 50 , 36 , 4 ] . Anticipating such regime shifts is critical for mitigating their potentially severe consequences. Early warning signals (EWS) have emerged as an important tool for detecting the loss of resilience that precedes these critical transitions [ 48 , 9 , 13 , 15 , 20 , 12 , 14 , 40 , 5 , 6 , 37 , 19 , 35 , 25 ] . One hallmark of reduced resilience is flickering: intermittent, noise-driven switching between coexisting stable states [ 15 , 34 , 2 , 29 ] . Flickering has been documented in diverse systems, including palaeoclimate records [ 47 , 46 , 17 , 18 ] , neuroscience [ 32 , 28 ] , hibernation physiology [ 42 ] , and lake ecosystem dynamics [ 49 ] . It often emerges before a bifurcation, when competing attractors become less stable and noise can drive rapid transitions between them [ 34 , 18 ] . This makes flickering a valuable EWS that can complement indicators based on critical slowing down. However, existing approaches to detecting flickering rely mainly on statistical metrics such as variance, autocorrelation, and skewness [ 15 ] . These methods can be sensitive to noise characteristics, data resolution, and confounding processes, limiting their robustness in practice [ 15 , 11 ] . In parallel, advances in artificial intelligence (AI) and deep learning (DL) have demonstrated considerable potential for identifying EWS of tipping points, particularly for detecting bifurcations in noisy, nonlinear systems [ 11 , 10 , 43 , 30 , 16 , 20 , 22 , 3 ] . Yet, despite its importance, the problem of detecting flickering using DL has not been addressed. This gap is significant: if DL models could reliably recognise flickering, they could expand our capacity to detect instability in a wider range of systems, especially when traditional statistical approaches fail. Here, we develop a convolutional–long short-term memory (CNN–LSTM) framework to detect flickering patterns in noisy time series. We train the model on synthetic data generated from simple polynomial functions with additive noise, and test its generalization to a suite of more complex stochastic dynamical systems. We further show that the deep learning model consistently outperforms variance-based generic early warning indicators, achieving higher discrimination between true flickering and null cases. Finally, we apply the trained model to empirical datasets, including dormouse body temperature measurements [ 42 ] and palaeoclimate records from the African Humid Period [ 47 ] . Our results show that deep learning can robustly identify flickering across both simulated and real-world systems, offering a new and broadly applicable tool for anticipating critical transitions. Box 1. What is flickering? Flickering refers to noise-driven, intermittent transitions between coexisting stable states in a dynamical system. As a tipping point approaches, these switches become more frequent, indicating reduced resilience. (A–C) Potential landscapes V(x) as the control parameter p changes, altering the number and stability of equilibria. In (B), two nearby equilibria coexist, and noise drives rapid switching between them. (D) Corresponding time series x(t) with dashed lines marking the equilibrium positions from (A–C). The frequent switching in (B) is characteristic of flickering as an early warning signal. 3 Deep learning model and dataset construction We implemented a convolutional–long short-term memory (CNN–LSTM) network to detect flickering patterns in noisy time series [ 45 , 1 , 27 ] . The CNN layers extract localized features across long temporal windows, while the LSTM layers capture temporal dependencies and preserve relevant sequence information through recurrent connections. The final dense layer outputs a binary classification indicating the presence or absence of flickering. This architecture was chosen for its proven ability to combine spatial feature extraction with long-range temporal modeling in time series analysis [ 21 , 20 , 10 , 22 , 11 , 3 ] . A schematic of the architecture is shown in Fig. 2 , and all layer specifications are provided in the Methods and Materials section. Synthetic time series were generated from a one-dimensional stochastic differential equation consisting of a linear term, a seventh-degree polynomial, and additive Gaussian noise. The control parameter p was varied to produce two categories: • Flickering series : p was linearly varied from an initial value p_{0} to the critical value p^{*} , producing intermittent switching between attractors near the transition. • Non-flickering series : p was fixed at p_{0} , keeping the system in a single attractor. The noise amplitude \sigma was tuned to balance state exploration and temporal structure preservation. Each time series was normalized using z -scores, and a rolling variance was computed in parallel to the raw signal. The final dataset comprised 2000 flickering and 2000 non-flickering series. Full equation parameters and coefficient ranges are detailed in the Methods and Materials. Each sample was represented as a multivariate time series with two input channels (raw signal and rolling variance). We trained a series of CNN–LSTM models with identical architecture and hyperparameters but different sequence lengths \{5000,6000,7000,8000,9000,10000\} . Each model was thus specialized to detect flickering patterns occurring over its respective temporal scale, with shorter windows capturing rapid local fluctuations and longer windows resolving extended pre-transition structures. Since genuine flickering requires a minimum temporal extent to manifest, overly short windows may miss the full pattern. All models achieved final training accuracies above 95\% . Training across multiple scales and ensembling the outputs provides robust detection across the full range of temporal signatures. Full methodological details, including preprocessing, interpolation, and ensemble aggregation, are provided in the Methods and Materials. Figure 2 : Deep learning model architecture for flickering pattern detection This figure shows the deep learning framework used to detect flickering patterns and provide early warnings of tipping points. Simulated time series with and without flickering ( p^{*} , p_{0} ) are pre-processed and used to train the model, which outputs a binary classification indicating the presence of flickering. Figure

Read the complete text