Project institution: Lancaster University
Project supervisor(s): Dr Michael Tso (Lancaster University), Prof Andrew Binley (Lancaster University), Prof Andrew Curtis (University of Edinburgh), Dr Elizabeth Cooper (Lancaster University) and Dr Paul Wilkinson (British Geological Survey)
Overview and Background
Many near-surface geohazards, e.g. landslides and failure of earthen dams, result from changes in flow and storage of subsurface fluids. Understanding the likelihood of such hazards is essential but often challenging because we lack of suitable subsurface monitoring technology. A range of geophysical sensors can provide a 3D time-lapse ‘movie’ below ground, but analysing such data is computationally intensive and typically done in isolation, making interpretation ambiguous—yet this uncertainty is often ignored. Computational limitations from traditional paradigms also limit the amount of data that can be used to improve the model, and the extent to which the solution space can be explored. This project aims to overcome these challenges by developing new approaches that couple multiphysics simulators, tailored for GPUs, to derive improved subsurface models with an assessment of uncertainty, thus significantly improving our ability to identify hazards and risk.
Methodology and Objectives
We will focus on application to two geohazards for which rich time-lapse geophysical datasets currently exist. Candidates include 3D electrical resistivity monitoring of dynamic moisture changes and mass movements on a hillslope or earthen dam and distributed acoustic sensing of unstable embankments. Each problem consists of determining the temporal evolution of an image of the subsurface based on many tens or hundreds of thousands of unknown parameters. Inversion of data from a single geophysical modality (e.g. electrical resistivity) can be challenging using conventional computation approaches. Furthermore, we often require near real-time execution of such inversions to assess the risk of possible hazards occurring, with adequate time for mitigating actions to take place. In this PhD we wish to solve the inverse problem coupled with a fluid flow simulator and evaluate all relevant sources of uncertainty (ambiguity) in the solution. This cannot be achieved with conventional computers and computational method.
We will develop geophysical and fluid flow simulators based on existing tools (including those developed by members of the supervisory team). Each simulator will require enhancement for application on GPUs. We will consider different strategies ranging from discretised grid-based partial differential equation solvers to surrogate modelling approaches.
We will then explore optimum strategies for coupling the geophysical and flow simulators, and any necessary data pre-processing strategies. Unlike conventional inversion of geophysical data, we will use the geophysical data (not the models) to constrain the parameters of the flow simulator. Methods for computation of predictive uncertainty in the coupled inversion will be explored, e.g. Markov chain Monte Carlo, data assimilation methods or variational inference. Insights derived from this research will benefit from GPU-acceleration of model coupling in general (e.g. coupling weather and land surface models, or different ecosystem services): this is extremely important because it is a core component in the analysis of the interdependencies within and between complex real-world environmental systems.
Teaser Project 1: Surrogate modelling of geophysical data
Simulations of many geophysical problems rely on spatio-temporal discretisation of an approximation to the governing equations. This can be computationally restrictive for many large-scale problems. Surrogate modelling (or emulators) offers a computationally attractive alternative, particularly when focussed on GPU-based architecture. A range of surrogate modelling approaches exist, e.g. those based on deep learning, which may be able to address significant nonlinearity of the problem. In this element of the PhD, we will explore surrogate modelling as an alternative approach to conventional geophysical simulation.
This project will involve the development of a GPU-accelerated simulator for simulations, inversions, and training surrogate models. Additionally, it will focus on creating and utilising GPU-based software frameworks to enhance the training of these surrogate models. Such advances would enable efficient coupled hydro-geophysical simulations for robust uncertainty quantification and long-term training scenarios (e.g., different CMIP scenarios).
Teaser Project 2: Estimation of information content
Data collection comes at a cost. This includes the provision, maintenance and operation of sensors, and the computational demands for interpretation. A widely overlooked question is, how much data do we need to develop a reliable model of the subsurface? In addition, given that we may have access to multiple geophysical modalities, how do we decide where to focus resources? And how will the uncertainty in our model reduce (or not) as we add more data? In this element of the project, we will investigate methods for assessing the value of information, such as through entropy measures, with a focus on developing techniques and GPU-based software. We will apply a suitable approach to the field study problems, to optimise data acquisition to best constrain the subsurface parameters of interest.
References and Further Reading
- Boyd, J. P., Chambers, J. E., Wilkinson, P. B., Meldrum, P. I., Bruce, E., & Binley, A. (2024). Coupled Hydrogeophysical Modeling to Constrain Unsaturated Soil Parameters for a Slow‐Moving Landslide. Water Resources Research, 60(10) (click here)
- Cooper, E. S., Dance, S. L., Garcia-Pintado, J., Nichols, N. K., & Smith, P. J. (2018). Observation impact, domain length and parameter estimation in data assimilation for flood forecasting. Environmental Modelling & Software, 104, 199–214 (click here)
- Lu, D., Liu, Y., Zhang, Z., Bao, F., & Zhang, G. (2024). A Diffusion‐Based Uncertainty Quantification Method to Advance E3SM Land Model Calibration. Journal of Geophysical Research: Machine Learning and Computation, 1(3) (click here)
- Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707 (click here)
- Strutz, D. & Curtis, A. (2024) Variational Bayesian experimental design for geophysical applications: seismic source location, amplitude versus offset inversion, and estimating CO2 saturations in a subsurface reservoir. Geophysical Journal International, 236(3), 1309–1331 (click here)
- Tso, C.-H. M., Johnson, T.C., Song, X., Chen, X., Kuras, O., Wilkinson, P., Uhlemann, S., Chambers, J. & Binley, A. (2020) Integrated hydrogeophysical modelling and data assimilation for geoelectrical leak detection, Journal of Contaminant Hydrology, 234 (click here)
- Zhang, X., Nawaz, M. A., Zhao, X., & Curtis, A. (2021). An introduction to variational inference in geophysical inverse problems. Advances in Geophysics, 62, 73–140 (click here)
- Information Theory in the Geosciences: A hub for the emerging community of practice for Information Theory in the Geosciences
