Reduced order modelling of land subsidence
Project Overview
This project develops a neural network to predict land subsidence at specific locations, where groundwater movements are the main driver of such subsidence. Machine learning is used to map soil profiles, which describe the current state of the soil, to corresponding land subsidence values at specific times. This process is facilitated through Bede, where Julia scripts are submitted to train the neural network.
What were the key results of your research project?
Soil profiles can be successfully mapped to land subsidence, but the neural networks developed in this project only capture general trends, and missseasonal variations. Deeper networks like LSTMs or DeepONets would be a promising direction for future work, in order to capture these seasonal variations, and become suitable for practical use.
Another improvement would be to relax the assumptions made when creating the soil profiles, which primarily arise from the use of Terzaghi's 1D consolidation equation. This would require additional geotechnical engineering expertise as well as considerable increases in data processing, but could enable much more precise, location specific soil profiles.
Non-dimensionalisation of the soil profiles could also further enhance the robustness of the training, especially with limited training data, alongside simply training with more labelled training data.
How do you feel you have benefitted from completing this internship and has it made you consider future career paths?
I feel much more confident in using HPC resources to conduct computationally intensive research, especially in the context of geotechnical engineering, where I can utilise such resources to solve problems that I have learnt over the past few years in my degree on a massive scale.
Naturally, I feel more confident in independent research, and working from advice given by supervisors and superiors without a clear answer. I had a similar experience when undertaking my dissertation earlier this year, so this was a sort of extension as I also had many directions and methods to try in this internship, with no singular clear answer.
This internship has therefore attracted me more towards scientific research, where HPC can be utilised to create practical solutions and methods based off existing ones, and I have definitely considered pursuing such a path more than before the start of the internship.