Noise and Vibration Challenges through the Physics-Enhanced Machine Learning lens

Physics-Enhanced Machine Learning (PEML) - also known as Scientific Machine Learning and grey-box modelling- is a natural evolution of Machine Learning (ML) for guiding high-consequence decision making in engineering applications by developing models which integrate information extracted from real-world data, physics-based models, and domain and expert knowledge. PEML addresses the limitations of data-driven only techniques by constraining the space of admissible solutions, and enabling robust, generalisable and interpretable models even with limited data. Multiple hybrid data-physics-knowledge approaches have been developed depending on the available information (physics and domain knowledge on the dynamical system vs informative data), engineering task (e.g., uncertainty quantification, identification of unknown physics/terms of a model or development of a time-evolving digital twin model), and complexity of the system/problem. During this lecture, PEML will be introduced and three broad groups of PEML approaches will be discussed: physics-guided, physics-encoded and physics- informed. Recent strategies developed for noise and vibration applications, including efficient uncertainty quantification of vibroacoustic models and identification of non-smooth nonlinearities such as frictional contacts, will be presented. Open challenges and opportunities for tackling challenges in noise and vibrations will be discussed.