The Synergy of Physics-Informed and Machine Learning Techniques for Model Personalization
Abstract
In many engineering and medical applications, model personalization is a critical step towards building effective digital twins that can predict and control complex physical systems. To achieve this, efficient and accurate methods are needed for solving forward, inverse or parameter/field estimation problems. These challenges can be addressed using physics-informed machine learning, which leverages physical knowledge encoded in parametric differential models to improve the predictive and generalization capabilities of data-driven models, overcoming the limitations imposed by data availability. This optimal synergy allows the reconstruction of distributed quantities of interest, such as (heterogeneous) material properties, velocity and pressure, mechanical displacement or electric potential fields, which may not be directly measurable.
In this talk, we present some numerical strategies we have developed to tackle the challenges of designing effective physics-informed machine learning techniques for model personalization. We provide examples related to cardiovascular applications.