Current students

Scientific machine learning tackles complex problems where physics plays an important role. Among its advantages is a greater robustness and generalisability of models. To improve adoption and trust, however, much has to be done regarding the explainability of its techniques. The present thesis aims to enhance physics-informed methodologies with explainability tools to improve their efficacy and trustworthiness, in the particular context of predictive maintenance of critical equipment.

The application of the aforementioned techniques concerns the monitoring of health and residual lifetime of mechanical components, such as rotation and transmission machinery, in relation to their usage profile.
Such use-cases would provide a testing ground for several Physics-Informed techniques, which incorporate both prior knowledge and data-driven insights to ensure robustness and explainability. In addition, a systematic approach must be designed to tackle the minimization of the need to resort to black-box inference components, and the quantification of uncertainty affecting the decisions taken by interpreting the model's outputs.
Both issues can be addressed via more explainable and physics-informed models, but also by employing models that can automatically evaluate their applicability conditions and inform end users on the confidence regarding their outputs, and following a Bayesian, anomaly-detection oriented paradigm whenever possible.