Numerical Model Construction with Closed Observables

Felix Dietrich, Gerta Köster, Hans-Joachim Bungartz

TL;DR

This paper introduces a novel method for constructing simplified, "macroscopic" numerical models from complex, "microscopic" simulations or real-world experiments. By leveraging time-lagged embedding and diffusion maps, they create "closed observables" that capture the essential dynamics with significantly reduced computational cost and memory requirements, enabling faster analysis and control.

What Problem Does This Paper Address?

Many real-world systems, from particle simulations to traffic flow, exhibit complex dynamics across multiple scales. Analyzing, optimizing, or controlling these systems is computationally prohibitive if one must rely solely on the detailed microscopic description. When macroscopic observations are available (e.g., from legacy simulators or experiments), a major challenge is to derive a simplified model that accurately represents these macroscopic variables without needing to simulate the entire microscopic system. Existing methods often struggle with black-box systems or require extensive computational resources.

Methods

The core of the paper lies in constructing "closed observables." This involves:

  • Time-Lagged Embedding (Takens' Theorem): This technique reconstructs the system's dynamics from a time series of observations. By using a sequence of past observations as coordinates, it creates a higher-dimensional "state space" that is dynamically closed, meaning it contains enough information to predict future states.
  • Diffusion Maps: This is a non-linear dimensionality reduction technique applied to the time-lagged embedding space. It effectively separates different time scales and noise, allowing for a significant reduction in the dimensionality of the state space while preserving essential dynamics. The resulting lower-dimensional coordinates are termed "closed observables."
  • Numerical Model Construction: The paper then uses these closed observables to numerically construct a new, simplified dynamical system. This involves interpolating the coordinate transformation (mapping from original states to closed observables), the dynamics of the closed observables, and an observer function that maps the closed observables back to the original system's output. This results in a time-autonomous model that can be simulated independently of the original microscopic system.

Datasets

Not discussed in the paper. The paper uses synthetic examples (a 2D ODE system and a transport-diffusion PDE) and a real-world example (train station passenger flow) generated by simulators or based on experimental data.

Key Insights and Ideas

  • "Closed Observables" as a Bridge: The concept of "closed observables" is key. These are variables derived from observations that are sufficient to predict future system behavior, effectively closing the dynamics at a macroscopic level.
  • Storage Efficiency is Paramount: A major motivation is the dramatic reduction in memory requirements. The paper demonstrates that their method can reduce storage needs by orders of magnitude, making complex system analysis feasible on less powerful hardware (e.g., a smartphone vs. a supercomputer).
  • Black-Box Compatibility: The approach is designed to work with systems where the underlying equations are unknown or too complex to analyze directly, relying solely on output data.
  • Separation of Concerns: The method separates the "offline" model construction phase from the "online" simulation and analysis phase, allowing for efficient use of computational resources.

Key Results

  • Significant Speedup: In the train station example, analysis and simulation with the constructed numerical model were orders of magnitude faster (seconds vs. hours) than using the original microscopic model.
  • Dramatic Storage Reduction: For a 2D ODE example, the method reduced the required storage from that of a supercomputer to that of a smartphone.
  • Accurate Approximation: Theorem 2 shows that the approximation error of the new model is primarily determined by the interpolation error of the coordinate transformation, dynamics, and observer function, and can be controlled by the accuracy of these interpolations.
  • General Applicability: The method is demonstrated on both low-dimensional ODEs and a PDE system with an infinite-dimensional observer (density field), showcasing its versatility.

Limitations and Future Work

  • Interpolation Error: The accuracy of the constructed model depends on the quality of the interpolants used for the coordinate transformation, dynamics, and observer. Theorem 2 quantifies this error, but practical implementation requires careful selection of interpolation methods.
  • Stochastic Systems: While the paper mentions stochastic systems, the current framework is primarily presented for deterministic systems. Extending it to handle stochasticity robustly is a future direction.
  • Analytic Forms: The paper suggests searching for analytic forms of closed observables, which could further improve efficiency and interpretability.
  • Network Optimization and Uncertainty Quantification: The authors propose applying this model construction technique to network optimization and uncertainty quantification, suggesting it can simplify complex systems for these analyses.

Summary updated

Abstract

Performing analysis, optimization and control using simulations of many-particle systems is computationally demanding when no macroscopic model for the dynamics of the variables of interest is available. In case observations on the macroscopic scale can only be produced via legacy simulator code or live experiments, finding a model for these macroscopic variables is challenging. In this paper, we employ time-lagged embedding theory to construct macroscopic numerical models from output data of a black box, such as a simulator or live experiments. Since the state space variables of the constructed, coarse model are dynamically closed and observable by an observation function, we call these variables closed observables. The approach is an online-offline procedure, as model construction from observation data is performed offline and the new model can then be used in an online phase, independent of the original. We illustrate the theoretical findings with numerical models constructed from time series of a two-dimensional ordinary differential equation system, and from the density evolution of a transport-diffusion system. Applicability is demonstrated in a real-world example, where passengers leave a train and the macroscopic model for the density flow onto the platform is constructed with our approach. If only the macroscopic variables are of interest, simulation runtimes with the numerical model are three orders of magnitude lower compared to simulations with the original fine scale model. We conclude with a brief discussion of possibilities of numerical model construction in systematic upscaling, network optimization and uncertainty quantification.
BibTeX
@article{dietrich2016numerical,
  author = {Felix Dietrich and Gerta Köster and Hans-Joachim Bungartz},
  title = {{Numerical Model Construction with Closed Observables}},
  journal = {SIAM J. Appl. Dyn. Syst.},
  year = {2016},
  url = {https://openreview.net/forum?id=EY9OQPpgyc},
}
Figure