Abstract:Complex nonlinear dynamics are a general characteristic of ecosystems and a challenge for scientists who seek to understand, model, and manage them. these complex dynamics can result from the interaction between endogenous feedbacks, external forcing, and noise (i.e., stochasticity) and can lead to gradual shifts, switching between alternative stable states, deterministic chaos, mirage correlations, and critical transitions. although the presence of nonlinear, state-dependent behavior means that it may not be possible able to predict the exact state of ecosystems 10–50 years from now, there remains a need for analytical tools that can address the nonlinear behaviors of real ecosystems. Such tools are important to produce accurate forecasts in the short-term, to understand key ecological mechanisms, and to enable adaptive management strategies that remain robust in response to such phenomena as climate change and invasive species. a fundamental problem with traditional modeling approaches is that they are based on parametric equations, which represent very specific hypotheses about system behavior that can be difficult to justify (Wood and thomas 1999). Moreover, these models often make simplifying assumptions (e.g., that systems are at equilibrium, different interactions are linearly separable) that do not line up with the known nonlinear reality. Consequently, these models can lack the flexibility to predict future behavior even when they are able to fit the observed data. For example, in nonlinear systems, correlations between variables may not be detectable even if they are dynamically coupled (fig. 1). thus, variables can appear correlated for the duration of a research study, but this correlation may vanish even though the dynamics have not changed. one alternative to conventional parametric models is the framework of empirical Dynamic Modeling (eDM), which does not assume any particular set of equations for the relationships between variables, but instead reconstructs behavior directly from time series data. eDM is based on takens’ theorem (takens 1981) and the method of time delay embedding (Packard et al. 1980), which enables it to describe dynamical relationships that are too complex or subtle to capture in a simple set of equations. a brief description of the methods can be found in this short animation: http://simplex. ucsd.edu/eDM_101_rMM.mov. in addition to the basic methodology of eDM, different techniques have been developed that extend the approach for specific applications: testing for the influ0.2 0.4 0.6 0.8 1 A B C

Using neural ordinary differential equations to predict complex ecological dynamics from population density data

Sparsity in Continuous-Depth Neural Networks

Divide And Conquer: Learning Chaotic Dynamical Systems With Multistep Penalty Neural Ordinary Differential Equations

A novel node approach combined with LSTM for short-term electricity load forecasting

HiDeS: a higher-order-derivative-supervised neural ordinary differential equation for multi-robot systems and opinion dynamics

Incorporating NODE with Pre-trained Neural Differential Operator for Learning Dynamics

Beyond Predictions in Neural ODEs: Identification and Interventions.

Predicting the Future in a nonlinear World

Learning the Delay Using Neural Delay Differential Equations

Neural Ordinary Differential Equations for Nonlinear System Identification

Scientific machine learning in ecological systems: A study on the predator-prey dynamics

On Second Order Behaviour in Augmented Neural ODEs

Enhancing Human-Machine Teaming for Medical Prognosis Through Neural Ordinary Differential Equations (NODEs)

Expressive architectures enhance interpretability of dynamics-based neural population models

An extended neural ordinary differential equation network with grey system and its applications

Electrical Load Forecasting Based on the Fusion of Multi-Scale Features Extracted by Using Neural Ordinary Differential Equation

Forecasting and predicting stochastic agent-based model data with biologically-informed neural networks

Forecasting with an N-dimensional Langevin equation and a neural-ordinary differential equation