Research

When learned models are useful — and when they fail

My research interests sit at the intersection of machine learning, physical systems, and engineering decision-making. I am especially interested in when learned models are useful, when they fail, and how they can support faster, more reliable engineering workflows.

01

Surrogate Modelling for Physical Systems

Surrogate models aim to approximate expensive simulations or physical processes with faster learned models. The challenge is not only prediction accuracy, but reliability, generalisation, and behaviour under changing physical regimes.

Open questions
  • When do learned surrogate models fail?
  • How can we evaluate them beyond one-step accuracy?
  • What makes a surrogate useful in engineering decision-making?
  • How do we balance speed, interpretability, and physical fidelity?
02

Structural Dynamics and Data-Driven Modelling

Physical structures generate complex dynamic responses. Machine learning can help identify patterns, infer states, or accelerate analysis, but models must respect the physical behaviour of the system.

Open questions
  • How can dynamic response data be used for inference?
  • Which signals matter most?
  • How robust are learned models under noise, damage, or distribution shift?
03

Scientific Machine Learning

Scientific machine learning sits between pure machine learning and physics-based modelling. The opportunity is not just better prediction, but better tools for discovery, design, and decision-making.

Open questions
  • Where does ML genuinely improve engineering workflows?
  • Where is physics still essential?
  • What does trustworthy ML look like in physical systems?
Research Principles
  1. 01Accuracy is not enough.
  2. 02Robustness matters.
  3. 03Physical meaning matters.
  4. 04Useful models must survive contact with real workflows.
  5. 05The best technical work should eventually improve a decision, workflow, or system.