Domain adaptation and active learning AI techniques in the context of regression, simulation and agnostic optimization of large industrial apparatuses and high energy physics experiments

Artificial intelligence (AI) and machine learning (ML) are increasingly at the head of technological advancements, driving innovation across a multitude of domains. In the field of physics, data analysis has always been a key aspect of scientific exploration, from the early days of astronomy and particle detection to the modern era of large-scale experiments and simulations. As experimental setups have become more complex and the volume of data has grown exponentially, traditional methods of analysis have evolved to incorporate more sophisticated computational techniques. The natural progression of these methods has led to the adoption of AI and ML, which offer powerful tools for uncovering patterns, making predictions, and automating processes in ways previously unattainable. Particularly important are the techniques that enable AI models to adapt to new and varied environments—known as Domain Adaptation—and to learn efficiently from limited data—through Active Learning. This thesis focuses on advancing these methodologies within two significant areas: the condition-based maintenance of large industrial apparatuses, specifically in collaboration with Nuovo Pignone Tecnologie s.r.l, part of Baker Hughes (BH), which is the funder of the PhD scholarship, and the rigorous Data Quality Monitoring (DQM) in high energy physics (HEP) experiments, as exemplified by the Compact Muon Solenoid (CMS) experiment at the Large Hadron Collider (LHC). The integration of ML into these domains not only enhances the ability to process and interpret vast amounts of data but also represents a significant step in the ongoing evolution of data-driven physics research.