Learning What We Cannot Measure in Engineering Systems

Modern engineering and physical systems generate increasingly large volumes of data through sensors, monitoring networks, and computational models. However, many of the quantities needed to understand their behavior, assess their condition, and inform decisions remain difficult or impossible to measure directly. Observations are often incomplete, labeled data are scarce, and measurements are influenced by changing operational and environmental conditions. Furthermore, different physical mechanisms can produce similar observable responses, making it challenging to distinguish between accurate predictions and reliable physical interpretations. These limitations raise fundamental questions about what can be learned from data, how knowledge can be transferred across systems, and how uncertainty should be characterized and propagated through predictive models.

In this seminar, Dr. Tronci will present research at the intersection of probabilistic inference and physics-based modeling, aimed at extracting meaningful information from incomplete and indirect observations of complex physical systems. Drawing on applications in civil infrastructure and offshore wind energy, Tronci will discuss how data-driven approaches can identify informative representations of system behavior, infer unmeasured physical quantities, and develop predictive models under varying operational conditions. Tronci will pay particular attention to Gaussian processes, Bayesian estimation, transfer learning, and hybrid modeling strategies that integrate computational simulations with experimental observations. These approaches will be illustrated through applications involving structural response prediction, load estimation, and digital twin development, highlighting their underlying assumptions, capabilities, and limitations.

The seminar will explore the broader challenge of establishing confidence in data-driven models when observations are limited and the underlying physical processes are only partially understood. Tronci will discuss how uncertainty and model assumptions influence the interpretation of monitoring data and the reliability of subsequent decisions.