Numerical Analysis and Scientific Computing Seminar | Jenmy Zhang, Probabilistic learning for model-form uncertainty quantification and digital twinning of real-world engineering systems

A digital twin is an accurate digital replica of a physical engineering system and its associated processes. Digital twins provide transparency into the physical twin's "as used" performance, improve the system's efficiency, and reduce downtime. Uncertainty in digital twins is inevitable due to modeling assumptions, randomness in the environment, and sensor noise etc., which must be quantified to ensure reliable predictions. In this seminar, we present a structured approach for building digital twins particularly for systems with unknown inputs. Our methodology integrates projection-based model order reduction, a rapid approach for identifying unknown inputs, and a non-parametric probabilistic learning approach for modeling and quantifying model-form uncertainty based on real-world observations. We use as a case study an indoor footbridge at the Autodesk Technology Center at Pier 9 in San Francisco. The input to the digital twin is a set of "what-if" or "actual" load parameters, the latter inferred based on measurements of the 6 strain gauges on the bridge. The output is a set of probabilistic predictions of the physical structure's response under the specified loads, including quantities and regions of interests that are not directly measured by sensors. The digital twin learns, from strain data collected over a 3-day window, the response and the associated uncertainty of a “healthy” structure. It can then serve as a baseline against which future changes in the structure can be reliably quantified and diagnosed. Our results demonstrate the viability of the approach, even with “gappy” sensor data due to partial corruption by unknown sources of interference, and they underscore the value of accurately modeling uncertainty to enhance the performance and reliability of digital twins in real-world engineering applications.