PhD Studentship: Bayesian Machine Learning and Sampling Methods for Geophysical Inversion using Large Scale Seismic and Reservoir Simulations, NERC GW4+ DTP PhD studentship for 2023 Entry. PhD in Earth and Environment Science

About the Partnership

This project is one of a number that are in competition for funding from the NERC Great Western Four+ Doctoral Training Partnership (GW4+ DTP).  The GW4+ DTP consists of the Great Western Four alliance of the University of Bath, University of Bristol, Cardiff University and the University of Exeter plus five Research Organisation partners:  British Antarctic Survey, British Geological Survey, Centre for Ecology and Hydrology,  the Natural History Museum and Plymouth Marine Laboratory.

For eligible successful applicants, the studentships comprises:

• An stipend for 3.5 years (currently £17,668 p.a. for 2022-23) in line with UK Research and Innovation rates
• Payment of university tuition fees;
• A research budget of £11,000 for an international conference, lab, field and research expenses;
• A training budget of £3,250 for specialist training courses and expenses

Project Background

In earth science for hydrocarbon and mineral exploration, determining subsurface and source properties from seismic traces are challenging tasks, commonly known as the full-waveform inversion (FWI) and seismic source inversion, respectively. Often, seismic data are buried under significant amounts of ambient noise and combined with uncertainties in the geological model which complicates the inversion process. Both the FWI and source inversion are important aspects of subsurface monitoring to constrain changing material properties and evolving stress-fields of large geological models.

Such inverse problems usually employ Monte Carlo simulation frameworks, requiring thousands of forward simulations on large complex geological models, which demand significant computing time and resource. This project will aim to accelerate this process using recent advances in Bayesian inference and machine learning, especially utilizing deep learning and deep Gaussian process models.

Stress accumulation and fluid flow movement monitoring in reservoir needs complex geophysical and petrophysical simulations using known velocity models, permeability, permittivity etc. Efficient management and processing of such large volumes of synthetic seismic and petrophysical data in a probabilistic geophysical inversion, imaging process, history matching and uncertainty quantification is an open challenge, with outcomes that will benefit both industrial and academic research.

Project Aims and Methods

This project will explore advanced signal/image processing, machine learning approaches, in particular, deep-learning and Bayesian inference for parameter/uncertainty estimation and probabilistic inversion of large-scale geological models fusing geophysical/seismic and petrophysical data. Here, the aim is to reduce computational time utilising the recent advancements in deep neural networks and deep Gaussian processes to approximate the physical data generation process, i.e. the seismic wave propagation and reservoir fluid flow simulations.

The concepts from seismic interferometry will also be used for turning highly noisy traces into useful interpretable signals using various correlation-based methods. Quantification of the uncertainties in such geophysical inverse problems in terms of both seismic source properties and unknown elastic geological models (density, compressional and shear wave velocity) and petrophysical parameters like permeability, porosity etc. is a complex problem. Geophysical inverse problems rely on the travel-time calculation between sources and receivers.

However, uncertainties in the velocity model can make these estimates highly erroneous. Alternatively, a full seismic wave based inversion can be attempted for improved imaging, albeit being computationally challenging. The project will also explore the inversion results of 3-component geophone recordings apart from pressure measurements by hydrophones in a marine environment. The traditional inversion or seismic imaging methods involve a series of heuristic filtering steps that can be more optimally selected using a deep machine learning based expert system.