Machine learning for non-linear aerospace systems: time-series diagnostics, model identification, and model order reduction

In aerospace engineering, accurately modeling complex nonlinear phenomena is crucial for system diagnostics, control, and optimization. Traditional physics-based models, while effective, often require extensive computational resources and may struggle to capture intricate, data-driven dependencies. Conversely, purely data-driven approaches, particularly machine learning (ML) techniques, offer strong pattern recognition capabilities but suffer from limited generalizability beyond the training data. This thesis explores a hybrid approach that integrates data-driven techniques with knowledge-based (physics-informed) models to enhance both interpretability and predictive accuracy in nonlinear aerospace systems. Three case studies are investigated: (1) Data-driven diagnostics for sloshing-structure interaction in aircraft wings, where ML techniques are applied to experimental data to extract coherent patterns and correlations; (2) Model identification for nonlinear panel flutter, leveraging sparse regression techniques to bridge the gap between system diagnostics and predictive modeling; and (3) Model order reduction for pressure collapse in cryogenic tanks, where a physics-based model is coupled with ML techniques to estimate parameters not readily obtainable from first principles. A key contribution of this work is the demonstration that hybrid modeling approaches improve generalization, mitigate extrapolation errors, and enhance experimental design efficiency. The integration of physics-based constraints with ML predictions allows for scalable, computationally efficient models applicable across different aerospace configurations. Experimental validation, particularly in the study of sloshing-induced pressure collapse, confirms the predictive accuracy of the proposed methodologies. The findings of this thesis contribute to the growing field of data-driven modeling in aerospace applications, offering a systematic framework for leveraging ML while preserving fundamental physical principles. By combining data-driven insights with expert knowledge, this research advances the reliability and applicability of predictive models for nonlinear aerospace systems.