Research

Thread 01

Ground-borne and building vibration

When a train passes, energy radiates into the soil, travels tens or hundreds of metres, and re-enters nearby buildings as perceptible vibration and re-radiated noise. Predicting that chain accurately, before construction, is the central problem.

The difficulty is that every link in the chain is uncertain. Soil is layered, heterogeneous, and rarely characterised in detail. Buildings have foundations whose stiffness is hard to pin down, and structural modes that interact with the incoming field. The wave field itself is three dimensional and radiates into an unbounded domain, so the usual finite domain shortcuts do not apply.

A large part of my work asks a deliberately practical question: which modelling assumptions actually change the prediction? Engineers routinely simplify by assuming rigid foundations, decoupling soil and structure, working in two dimensional sections, or treating soil as linear. Each simplification saves enormous computational effort. My research quantifies what those choices cost in accuracy, so that practitioners can simplify deliberately rather than by habit.

I have applied this to vibration from railways, from simultaneous road and rail traffic at crossings, and from human activity inside buildings, where the source is close to the receiver and the modelling assumptions matter most.

Thread 02

Numerical methods for soil-structure interaction

Better predictions need better solvers. Part of our group's work looks at coupled formulations that handle unbounded soil domains and several interacting structures without the cost of brute force discretisation.

Finite elements work well inside structures but are less suited to unbounded soil, where they need artificial boundaries and large meshes. Boundary type and singular boundary methods handle the radiating domain more naturally, though they are harder to apply to structural detail. Coupling the two aims to draw on the strengths of each, and the coupling itself is where most of the technical difficulty sits.

We have contributed to a coupled FEM-SBM methodology for structure-soil-structure interaction: the case where several structures sit close enough that each one scatters the field arriving at its neighbours. This effect is often neglected in practice, and there appear to be cases where neglecting it is not safe.

Thread 03

Machine learning for vibration analysis

Some questions in this field are not limited by physics but by compute. Where that is true, learned surrogates change what is answerable.

A single high fidelity soil-structure simulation can take hours. That is acceptable for a final check and useless for the questions engineers actually want answered. How does the response vary across the plausible range of soil properties? Which parameters dominate? Given measured vibration at the surface, what is the most likely subsurface profile?

Each of those requires thousands of model evaluations. I use machine learning to build surrogates trained on a manageable number of full simulations. Those surrogates then answer the parametric and inverse questions in seconds, with the physics model kept in the loop for verification rather than replaced by the surrogate.

Thread 04

Uncertainty quantification

Almost nothing in this problem is known exactly. Treating a vibration prediction as a single number hides that fact. Treating it as a distribution makes the honest answer visible and, in practice, far more useful.

Uncertainty enters at every stage. Ground properties are inferred from a small number of boreholes and geophysical tests, then extrapolated across a whole site. The building carries its own unknowns: material properties, damping, foundation and connection stiffnesses, few of which are ever measured on the structure actually being assessed. The source varies with vehicle, speed, and track condition. On top of all of that sits model uncertainty, the error introduced by the modelling choices themselves, which can be larger than the scatter in any single input.

I work on propagating these uncertainties through the model so that the output is a range with a stated confidence rather than a single fragile figure, and on the sensitivity analysis that shows which inputs are actually driving that range. Both pair naturally with the surrogate models above, because credible uncertainty quantification needs many model evaluations and that is exactly what a fast surrogate provides.

It also changes the conversation with clients and authorities. A prediction with quantified confidence supports a proportionate decision about mitigation, instead of forcing a choice between over engineering and hoping for the best.