A0372
Title: Comparative analysis of model discrepancy treatment: Calibration versus scientific machine learning
Authors: Victoria Volodina - University of Exeter (United Kingdom) [presenting]
Abstract: In healthcare and biological systems, mathematical models are increasingly used to understand complex biological processes. However, many of these models suffer from model inadequacy, posing a significant challenge to their use in clinical decision-making. Two main methods are reviewed to deal with model errors. Within the uncertainty quantification (UQ) community, model error is considered during calibration, where the observation is expressed as the sum of three terms: the simulator output at the true values of the calibration parameters, the model discrepancy, and the observation error. In calibration, a mathematical model is treated as a "black-box" system with the main objective of learning the values of calibration parameters. An alternative approach is to construct a hybrid "grey-box" model by filling in the incomplete parts of the computational model with a non-parametric model. The non-parametric model is used to learn the missing processes by comparing the available observations with the computational model output. To provide interpretability, the outputs of the non-parametric model are then regressed back down to symbolic form to learn the missing terms from the model using symbolic regression. These two methods are compared, and their performance is illustrated in an application to Siggaard-Andersen oxygen status algorithm.
