Machine learning methods trained on simple models can predict critical transitions in complex natural systems

Abstract

Forecasting sudden changes in complex systems is a critical but challenging task, with previously developed methods varying widely in their reliability. Here we develop a novel detection method, using simple theoretical models to train a deep neural network to detect critical transitions—the Early Warning Signal Network (EWSNet). We then demonstrate that this network, trained on simulated data, can reliably predict observed real-world transitions in systems ranging from rapid climatic change to the collapse of ecological populations. Importantly, our model appears to capture latent properties in time series missed by previous warning signals approaches, allowing us to not only detect if a transition is approaching, but critically whether the collapse will be catastrophic or non-catastrophic. These novel properties mean EWSNet has the potential to serve as an indicator of transitions across a broad spectrum of complex systems, without requiring information on the structure of the system being monitored. Our work highlights the practicality of deep learning for addressing further questions pertaining to ecosystem collapse and has much broader management implications.

EWSmethods: an R package to forecast tipping points at the community level using early warning signals, resilience measures, and machine learning models

“…The resulting surrogates are therefore representative of the assessed non-cyclical time series. Alternatively, a fourth option is applicable for EWSNet following the original authors' suggestions, where a probability larger than 0.33 (the chance that all outcomes are equally likely) is indicative of an approaching transition (Deb et al 2022).…”

“…Multivariate forms of EWSs (Weinans et al 2019, Lever et al 2020, Medeiros et al 2022) and deep learning models (Bury et al 2021, Deb et al 2022) are of particular interest as they appear superior tools to the univariate signals described above. Multivariate approaches exploit information from multiple measurements of a shared system (e.g.…”

“…Our identification of the phenomenon stems from linear stability analysis (LSA) of mathematical models (Ludwig et al 1978, Scheffer et al 2009), but machine learning tools can identify other phenomenological features not detected from LSA. For example, machine learning models trained upon transitioning data outperform equivalent models trained upon the EWSs of the same transitioning data (Deb et al 2022). This is indicative of alternative features being more informative than CSD to warrant general usage, although the 'black-box' nature of the approach limits its accountability (Enni and Herrie 2021).…”

“…This is indicative of alternative features being more informative than CSD to warrant general usage, although the ‘black‐box' nature of the approach limits its accountability (Enni and Herrie 2021). This being said, multiple machine learning models are now available for transitioning systems that improve the transparency of predictions by training on simple mathematical models associated with LSA (Bury et al 2021, Deb et al 2022). These models can consequently build upon our foundational knowledge of tipping points by taking advantage of the biases inherent in their training.…”

Tipping points in spatial ecosystems driven by correlated noise

“…However, robust indicators that forewarn such transitions are lacking, and results confirm that generic S-EWSs have had mixed success and their reliability reduced further with more correlated noise [38]. However, despite the challenges, recent work has sought to improve early warning signals using deep learning techniques [39]. While such methods are propitious, it requires further fine-tuning before such models serve as universal indicators of critical transitions.…”