Reconstructing dynamics from data: Algorithms for the prediction, estimation, and correction of chaotic systems
Date
2025
Authors
McGowan, Sean Peter
Editors
Advisors
Balasuriya, Sanjeeva
Robertson, Will
Robertson, Will
Journal Title
Journal ISSN
Volume Title
Type:
Thesis
Citation
Statement of Responsibility
Conference Name
Abstract
Methods for accurately predicting and estimating the states of chaotic systems are of considerable interest to many disciplines. The interacting effects of sensitivity to initial conditions, imperfect modelling, and noisy observations cause challenges for their application. As sensors become ubiquitous, approaching this problem from a data-driven perspective may offer improved accuracy compared to traditional modelbased forecasting. It is still of great importance that these data-driven methods provide interpretable and meaningful analysis of the system of interest, as is the goal of standard modelling. Motivated by the imperfections of both models and data, this thesis explores methods of reconstructing dynamical systems from observational data to empower novel techniques for prediction, estimation, and correction of chaotic systems. The algorithms developed in this work rely on Takens embeddings for the reconstruction of attractors from partially observed chaotic data, as well as ergodic hypotheses for well-defined function approximation of dynamic maps and flows. From these premises, we develop methods to reconstruct dynamics for their analysis, prediction, and estimation. Using an optimal control approach, a technique to extract rules governing the evolution of data is presented which allows the prediction, uncertainty quantification, and Lyapunov analysis of partially observed systems in settings of high noise and limited data. We extend these techniques to a data assimilation framework, where new observations of the system may be used to update the prediction in a model-free manner which shows promise in settings where models are imperfect or not available. Finally, these ideas are used to correct the bias between observations and model output to leverage the mechanistic understandability of knowledge-based models with the predictive power of data-driven methods in a way that provides interpretability of coherent unmodelled dynamics. These approaches are validated on simulated chaotic systems as well as empirical datasets of meteorological variables. This research suggests that data-driven techniques may outperform model-based methods with tasks of prediction and estimation in chaotic, noisy, and partially observed settings particularly when available models are imperfect. They also may assist in offering new methods of extracting interpretation and mechanistic understanding to improve existing models with missing dynamics.
School/Discipline
School of Computer and Mathematical Sciences
Dissertation Note
Thesis (Ph.D.) -- University of Adelaide, School of Computer and Mathematical Sciences, 2025
Provenance
This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legals