CFE-CMStatistics 2024: Start Registration
View Submission - CFECMStatistics2024
A1344
Title: Quantifying the value of domain knowledge in physics-informed machine learning Authors:  Aidan Hughes - The University of Sheffield (United Kingdom) [presenting]
Elizabeth Cross - University of Sheffield (United Kingdom)
Keith Worden - The University of Sheffield (United Kingdom)
Sam Gibson - The University of Sheffield (United Kingdom)
Timothy Rogers - The University of Sheffield (United Kingdom)
Matthew Jones - University of Sheffield (United Kingdom)
Abstract: Predictive models are used to support decision-making in a variety of tasks within the field of engineering, including design, control, and maintenance planning. Recent research into physics-informed machine learning (or grey-box modelling) has sought to overcome the limitations of models based solely on physics or data. Physics-informed machine learners possess enhanced predictive capabilities, retaining the flexibility of data-based models while also being able to leverage the domain knowledge encoded in physical equations and constraints for regions of the input space where data are scarce. Examining the different modelling approaches from a decision-theoretic perspective is desirable to highlight the advantages of physics-informed machine learning models to end-users and to accelerate the adoption of such models into engineering decision processes. To this end, an approach is presented for quantifying the value of information associated with the domain knowledge present in physics-informed machine learning models. The value of information is a concept from decision theory that captures the difference in the maximum (expected) utilities achieved for a given decision process when solved with and without said information. The approach is demonstrated using a simulated decision problem framed around a numerical case study pertaining to the operation and maintenance of a structure.