Georgia Institute of Technology
Physics-Informed Neural Network Approaches to Stability Learning and Control Applications
Abstract
dc:description.abstractStability analysis of dynamical systems involves both qualitative and quantitative aspects. While qualitative stability determines whether an equilibrium is stable, quantitative stability characterizes the size of the region of attraction (ROA), namely the set of initial conditions from which system trajectories converge to the equilibrium. Accurate characterization of stability and its ROA is fundamental to safety-critical control design, verification, and certification. However, classical analytical tools, such as Lyapunov-based constructions, sum of squares (SOS) optimization, and partial differential equation (PDE)-based approaches, often become conservative, computationally expensive, or intractable when applied to nonlinear, high-dimensional, or uncertain systems. Recent advances in neural networks offer a promising alternative. As universal function approximators, neural networks can parameterize Lyapunov functions or approximate solutions to stability-related PDEs, thereby enabling data-driven estimation of stability certificates and regions of attraction. By embedding Lyapunov conditions or PDE constraints into the training objective, neural networks can provide structured approximations with theoretical guarantees. Despite their expressive power, neural network-based approaches face several challenges. They may suffer from overfitting, lack of robustness, or convergence to degraded solutions, especially when training data are sparse or unevenly distributed across the state space. Furthermore, in the presence of disturbances or model uncertainty, learning the ROA becomes significantly more challenging due to worst-case considerations and the need to enforce robustness constraints. Standard training formulations often fail to capture these robustness requirements, leading to overly optimistic or unstable estimations. To address these limitations, this dissertation develops enhanced neural stability-learning frameworks. For Lyapunov function construction, we introduce an augmented learning strategy, which replaces uniform sampling with verifier-generated informative samples to improve efficiency and reliability. For robust region of attraction (RROA), we propose physics-informed neural network (PINN)-based framework combined with policy iteration and rollout-generated anchor data, enabling stable approximation of the viscosity solution of the generalized Zubov equation under disturbance effects. The proposed methods are validated on complex nonlinear systems, including high-dimensional models, multi-equilibrium opinion dynamics, and disturbance-affected control systems. Experimental results demonstrate improved accuracy, convergence behavior, robustness, and scalability compared to existing approaches, highlighting the effectiveness of the proposed frameworks for practical stability analysis in challenging dynamical settings.
Degree
thesis:*- Name thesis:degree_name
- Electrical and Computer Engineering, PhD
- Department thesis:degree_department
- Electrical and Computer Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wang, Junkai
- Advisor dc:contributor.advisor
-
- Wardi, Yorai Y.
- Committee members dc:contributor.committeemember
-
- Zhang, Fumin
- Lee, Kok-Meng
- Tucker, Maegan
- Hale, Matthew
- ORCID person:identifier.orcid
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81711