Equation Free Modeling of Complex Dynamical Systems

55 concepts45 papers

For centuries, understanding a dynamical system meant writing down its governing differential equations from first principles. If you knew the physical forces at play, you could write down Newton's laws; if you understood fluid properties, you could rely on the Navier-Stokes equations.

But what happens when you observe a system whose underlying rules are completely unknown, too intricate to write down, or constantly changing? Consider the fluctuating prices in a financial market, the shifting gene expression levels in a developing cell, or the chaotic turbulence in a plasma discharge. Traditional system identification tries to fit data into predefined mathematical templates. The emerging paradigm of equation-free, data-driven modeling flips this script, extracting the underlying operators, topological structures, and predictive rules directly from raw observations.

Linearizing Chaos via Operator Theory

How can we predict highly nonlinear, chaotic systems without knowing their equations? The secret lies in a profound shift of perspective: trading a finite-dimensional nonlinear problem for an infinite-dimensional linear one. This is the core of Koopman operator theory. Instead of tracking how a system's state evolves nonlinearly, we track how functions of that state evolve linearly.

As highlighted by Frank Noe et al., transfer operators like the Koopman operator provide a unified mathematical lens for data-driven model reduction across fluid mechanics and molecular dynamics. A practical manifestation of this is Dynamic Mode Decomposition (DMD). While originally designed for fluid flows, the work of Jordan Mann demonstrates that DMD can extract predictive signals from raw financial trading data without assuming any underlying economic laws.

Nonlinear State Space (Hard)  ───►  Infinite-Dimensional Linear Space (Easy)
      x_t ──► f(x_t)                      ψ(x_t) ──► K ψ(x_t)

But how do we handle the "infinite-dimensional" requirement in practice? Steven L. Brunton et al. introduced the HAVOK (Hankel Alternative View of Koopman) framework, which shows that complex chaotic dynamics can be accurately represented as a simple linear system driven by intermittent, localized "kicks." This allows researchers to anticipate sudden, catastrophic shifts before they happen.

Furthermore, the utility of Koopman modes extends to real-time security; Sai Pushpak Nandanoori et al. leverage them to identify cyber-attacks on smart power grids purely from data streams, completely bypassing the need for fixed network topology models. When individual tracking is impossible, Alex Townsend et al. extend this by introducing the Distributional Koopman Operator, which allows modeling complex stochastic dynamics using only aggregate snapshot data.

Learning Continuous Propagators and Neural Fields

If we don't have explicit differential equations, how do we step forward in time? One approach is to learn a continuous-time vector field or a discrete "propagator" directly from data using machine learning.

To capture the continuous nature of physical systems without a predefined mathematical structure, Markus Heinonen et al. combined the flexibility of Gaussian Processes with the structural constraints of Ordinary Differential Equations (ODEs). This allows a model to learn the "rules of motion" from noisy, irregularly sampled data. To handle spatiotemporal systems, Suryanarayana Maddu developed STENCIL-NET, which bypasses partial differential equations entirely by learning a data-driven discrete propagator that mimics classic numerical schemes. Because it learns the underlying computational stencil rather than a specific solution, it generalizes remarkably well to coarser grids and larger spatial domains.

When dealing with long-term predictions, purely data-driven models often suffer from error accumulation, causing trajectories to drift physically. To solve this, Junfeng CHEN et al. introduced Deep-OSG, a framework that explicitly embeds the algebraic semigroup property into the neural network architecture. This ensures that stepping forward by time $t_1$ and then $t_2$ is identical to stepping forward by $t_1 + t_2$, preserving long-term stability.

For systems undergoing qualitative changes, the TRENDy framework by Matt Ricci et al. pairs multiscale filtering with Neural ODEs to successfully forecast complex bifurcations—such as shifting patterns on biological skin—directly from messy observational data.

Uncovering Geometry and Topological Invariants

Often, the goal of dynamical systems analysis is not just predicting the next state, but understanding the qualitative landscape: Where are the stable fixed points? Is the system oscillating? Is it chaotic?

When the governing equations are missing, we must extract these geometric properties directly from data. For instance, Bo Lin et al. developed a machine learning framework to compute "quasipotentials"—the non-equilibrium equivalent of an energy landscape—directly from observed trajectory data, bypassing the curse of dimensionality that cripples grid-based solvers. Similarly, Tal Shnitzer et al. combined manifold learning (Diffusion Maps) with control theory (Contracting Observers) to prove that the intrinsic coordinates of a nonlinear system can be recovered directly from the local covariance of observed data.

To bridge the gap between messy, high-dimensional biological observations and clean theoretical models, Noa Moriel et al. proposed Smooth Prototype Equivalences (SPE). By training invertible neural networks to map experimental data to simple "prototype" models, they can localize stable oscillations and fixed points in sparse, noisy datasets. In another work, Noa Moriel et al. developed a deep learning framework to extract the "topological DNA" or invariants of a system, allowing them to classify qualitative shifts across entirely different physical and biological domains from noisy snapshots.

Even without full phase space reconstruction, simple diagnostics can be highly effective: Arun Iyengar et al. demonstrated that the "0-1 test" can detect deterministic chaos directly from raw, high-frequency plasma data, completely bypassing the need for complex embedding parameter estimations. Similarly, in medical diagnostics, Jinling Yan et al. developed a model-free method using single-sample Jensen-Shannon Divergence to detect early warning signals of critical disease transitions directly from empirical gene expression data.

Model-Free Control, Filtering, and Inference

Extracting the dynamics is only half the battle; the ultimate test of an equation-free model is whether we can use it to infer hidden states or safely control a system.

In state estimation, classical filtering requires explicit transition matrices. To break this bottleneck, Cesare Alippi et al. introduced Graph Kalman Filters, combining Spatio-Temporal Graph Neural Networks with recursive Kalman filtering to estimate states on dynamic graphs where the underlying physics are unknown. For partially observable environments, Giacomo Arcieri et al. developed Deep Belief Markov Models to autonomously learn transition and observation dynamics for POMDP inference directly from raw action-observation data.

Raw Data Streams ──► Kernel/Neural Embeddings ──► Model-Free Control Action

When it comes to control and safety verification, Adam Thorpe et al. leveraged kernel distribution embeddings in reproducing kernel Hilbert spaces (RKHS) to perform stochastic reachability analysis. By learning uncertainties and dynamics directly from sampled trajectories, their approach eliminates the need for explicit transition kernels. This data-driven reachability framework was expanded into the SOCKS toolbox by Adam Thorpe et al. to handle stochastic optimal control with arbitrary cost functions.

These model-free control paradigms are proving vital in real-world engineering and economics:

  • Autonomous Systems: Gábor Rödönyi et al. utilized a neural network-based subspace-encoder method to capture the complex, proprietary steering dynamics of autonomous vehicles.
  • Industrial Processes: Jiaqi Meng et al. deployed an RNN-LSTM predictive model to successfully execute Model Predictive Control for a complex corn-to-sugar process without a first-principles model.
  • Macroeconomics: Where transition dynamics of market shocks are notoriously elusive, Ruitu Xu et al. demonstrated that reinforcement learning can infer stable market equilibria directly from observed microeconomic data.

By moving away from rigid symbolic equations and toward flexible, operator-based representations, data-driven system identification bridges the gap between pure data science and classical dynamical systems theory.

Key Foundational Work

Gottwald Melborune (0–1) test for chaos in a plasma (2012) — contributed 2 concept elements to this direction.
Nonparametric Uncertainty Quantification for Stochastic Gradient Flows (2015) — contributed 1 concept elements to this direction.
Dynamic Mode Decomposition for Financial Trading Strategies (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_data_driven_system_identification_87d5d948,
  title = {data-driven system identification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-driven-system-identification/rdr_87d5d948cb03438d9058f81ada284956},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_data_driven_system_identification_87d5d948_v1,
  title = {data-driven system identification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-driven-system-identification/rdr_87d5d948cb03438d9058f81ada284956/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake