Trusted Predictive Modelling for Engineering Systems with Large Scale Data

Recent advances in Machine Learning (ML) combined with increasing computing power, storage capacity, and sensor technology have engineering industries scrambling to take advantage of the insights from big data. Digital Twins (DTs) exemplify this shift, increasingly incorporating data-driven ML models to support decision making, yielding cost savings. However, the utility of a DT depends on the trust placed in their predictive output, which can be driven by accuracy and uncertainty quantification. Trust in DTs can be inhibited by imperfections in data arising from poor coverage, the presence of noise and unmeasured system disturbances.

Physics-informed Machine Learning (PIML) is an emerging field that seeks to address data weaknesses through the embedding of physical knowledge about a process into ML model training. Multi-Fidelity (MF) modelling is another field that seeks to optimally combine information sources with different fidelities to improve the accuracy upon estimates about a process, which may be utilised to provide a source of physics via engineering models. There is great potential for the methods contained in these fields to address key trust challenges in DTs; however, research gaps remain. Using MF modelling, within a physics-informed paradigm, overcomes the challenges of complex system representation in many PIML methodologies but has a lack of principled schemes for trading off between data of different fidelity. Furthermore, the current MF methods with uncertainty quantification have high computational cost which scale with data size.

This thesis addresses these challenges. Insights into the limitation of Physics-Informed Neural Networks (PINNs) have motivated the development of two novel Bayesian MF modelling frameworks. The frameworks enable engineers to incorporate prior engineering knowledge about the datasets in a mathematically principled manner, yielding accurate predictive model outputs with uncertainty quantification. The frameworks are implemented on probabilistic Neural Networks (NNs) and demonstrate cost effective inference and improved scalability over existing methods.