Neural Network Modelling For Solving Inverse Problems In Engineering
Supervisor: Dr Andrew Liew
Joint Supervisor: Dr Danny Smyl
Neural networks are a powerful modelling tool for efficiently describing and predicting the behaviour of complex systems, particularly those that involve many input parameters or are calibrated with huge quantities of data. This trait, along with their accuracy (which can improve with time and with access to more data) has seen them applied to the fields of finance, speech/hand-writing recognition, driver-less cars, the digital realm, robotics, weather modelling and the medical field, to name just a few. However, their application in the civil and structural engineering field remains in its infancy, possibly due to the accessibility of large quantities of data, the difficulty in modelling real-world materials, structures, uncertainty and behaviour, and the scale and nature of relevant construction applications.
This research project will capitalise on the power, accuracy and speed of neural networks to model real-world engineering problems, both in a typical forward manner, where data input is passed to a model and a prediction is made, and also in the inverse direction, where we ask what are the input parameter values that lead to a given result. Problems to solve include: the construction sequencing of structures based on uncertainties, tolerances and the current as-built state, structural engineering design and optimisation of what would involve complex finite element models, using neural networks to predict or determine the initial states that caused collapse in ultimate limit design, or the use in structural health monitoring where the models can be supplementary to real-time and historical monitoring data for the prediction of corrosion, cracking, fatigue or recommending the need for remedial works. The PhD will utilise various neural network models generated from libraries such as TensorFlow and Keras, and utilise custom and commercial finite element software for the simulation of real engineering problems and generation of data. Neural networks can be based around traditional dense formulations, convolution neural networks for image data, and also experimenting with graph neural networks and generative adversarial networks.
This position is for a PhD student working full-time, after the successful award of a departmental EPSRC Doctoral Training Partnerships (DTP), or through a self-funded / scholarship placement. To apply please send a two-page CV and covering letter to email@example.com.
This project is NOT FUNDED, although Departmental/University scholarships are available for applicants who can demonstrate strong evidence of research potential.
The successful applicant is likely to have a first degree in an engineering discipline (civil, structural, mechanical etc) or in computer science. He/she will also have sound mathematical and computer programming skills.