COURSE 4: From white to grey: fusing physics with data for DTs

Course description:

When using Machine Learning (ML) for Digital Twins (DTs), an alternative mindset is required to build sensible representations from data. Unlike other ML applications (e.g. large language models) the datasets are typically small and curated - collected via sensing and simulation, rather than scraped from the internet. The limited variance of training data typically renders learning by ‘brute force’ infeasible (and unnecessary). Instead, we must encode domain-specific knowledge within ML algorithms, to enforce structure and constrain the space of possible models.

This course introduces ML for physical systems via statistical learning. Through case studies, we will fuse scientific knowledge with insights from data, using semiparametric models and constrained Gaussian process regression.

Objectives and target groups:

The course serves as an introduction to the framework of grey box modelling with physics-informed ML. Concepts are introduced via two practical examples, familiar to a science and engineering audience including (i) materials tests and (ii) wind turbine power prediction. The first uses more conventional semi-parametric modelling, while the second uses Gaussian process regression with constraints. Rather than focusing on analytical solutions to inference problems, each example works around the general framework of gradient ascent for various models (considering applications in auto-grad software). As such, the material covers

Target groups:

• Academics and practitioners with a more conventional engineering background
• Post-graduate students moving into DT research (and more widely) with applied ML
• Those interested in interpretable, UQ-based approaches to ML

Course Outline:

Course Lecturers:

Lawrence Bull

University of Glasgow | Research Fellow | Statistics & Data Analytics | School of Mathematics and Statistics
Scotland