On decision-making for adaptive models combining physics and data

Digital twins are an exciting new technology that aims to optimally combine data- and physics-based models to enchance predictive capabilities and aid in decision-making for high-value and safety-critical assets. A key aspiration for a digital twin is to dynamically adapt to previously unseen structural, operational, and environmental conditions. Given this goal, digital twins are required to recognise when poor predictive performance occurs. This requirement leads to the question; what actions may be taken to improve the predictive performance of a model combining physics and data? The current paper aims to highlight and discuss the decision process to be undertaken by an agent tasked with maintaining the predictive performance via the following courses of action: do nothing, parameter calibration, (re)learning of a data-based component and incorporation of new physics into the model. As a case study, a grey-box formulation with a GP-NARX model is used to predict the outputs of an asymmetric Duffing oscillator. This study is then used to motivate discussions around factors influencing the decision process and the benefits and challenges of a risk-based decision-making approach.