Project highlights

  • Develop novel data processing methods for classifying and generating palaeo-landscapes;
  • Make use of land surface evolution modelling and 3D seismic data;
  • Employ stochastic techniques for improving the accuracy of automated classification

Overview

Understanding the nature of the subsurface is crucial for many applications in the earth sciences. This requires knowledge of the evolution of sedimentary systems and the processes that acted to shape the land surface. The aim of this project is to improve understanding of the characteristics of such surfaces so that fundamental properties of sedimentary basins, such as their porosity, can be better understood. The primary sources of data will be modern relief, the output from palaeo-landscape evolution modelling and pre-existing 3D seismic data.

Basic scientific questions concern the statistical properties of land surfaces as a function of uplift, resistivity and erosive rates. This requires means to characterise such surfaces (e.g. Keylock et al., 2020). Given such techniques, follow-on questions concern the statistical significance of the values obtained and the ability of machine learning methods to depict and classify such surfaces correctly. Our gradual multifractal reconstruction framework (Keylock, 2017, 2019) is well-suited to such hypothesis testing. The figure shows an original DEM (η = 1) and three artificial surfaces that are maximally randomised with regards to the form of the algorithm (η = 0). Significant differences exist between the synthetic and real surfaces until at least (η = 0.6) for various terrain metrics. Thus, we can determine the relative complexity of different landscapes and also generate training datasets for machine learning algorithms to enhance the fidelity of automatic classification methods. The outputs from palaeo-landscape evolution models may be analysed in the way described, but there is also the potential within the project to use related statistical methods to analyse 3D seismic datasets to recover information on the properties of past erosional surfaces.

This project would suit a graduate from a geophysics or quantitative earth science background with experience with computational methods (numerical modelling, machine learning, data analytics), or a data scientist who is interested in gaining expertise in earth sciences. The emphases within the project will be tailored towards the academic strengths of the applicant.

Host

Loughborough University

Theme

  • Dynamic Earth

Supervisors

Project investigator

  • Chris Keylock, School of Architecture, Building and Civil Engineering, Loughborough University, Loughborough, Leicestershire LE11 3TU, [email protected]

Co-investigators

How to apply

Methodology

The project will have an innovative methodological basis as it will employ methods for characterising terrain, and for generating random topographic surfaces that have been developed recently (see references). These will be applied to the outputs from numerical models and 3D seismic data to determine the properties of the topographic surfaces. The final part of the project will relate to the efficacy of machine learning methods for classifying surfaces, and the extent to which this can be enhanced through the use of generating synthetic surfaces as part of the training process. I.e. given artificial surfaces that are not significantly different to the original based on measures of the terrain characteristics, what was a point in a multivariate space becomes an ellipsoid (or other shaped object), permitting distance measures to be formulated more effectively than in traditional data classification techniques.

Training and skills

Training in programming in a variety of languages, use of high-performance computing facilities etc is available centrally at Loughborough. The Loughborough supervisor developed the statistical methods to be employed in the project and will provide training on these. Landmark will provide training on the use of their palaeo-landscape modelling tools, on interrogation of the seismic data, and on the use of machine learning algorithms.

Partners and collaboration

The partner for this project is Halliburton Landmark based in Abingdon, Oxfordshire. The team have extensive experience with handling and interpreting large geophysical datasets, coding numerical process models and with associated data analytics tasks for automated classification. There may be the potential for this project, or a subsequent one building on the research, to be submitted as a CASE partnership. However, in the short time between initial contact and proposal submission, we have not been able to ascertain if this is possible given the need for budgetary approval within Halliburton.

Further details

For further information about this project, please contact Prof Chris Keylock ([email protected]; http://www.chriskeylock.net). For more information about CENTA and the application process, please visit the CENTA website: www.centa.ac.uk. For further enquiries about the application process, please contact the School of Architecture, Building and Civil Engineering ([email protected]). Please quote LU11_CENTA when completing the application form: http://www.lboro.ac.uk/study/apply/research/.

Possible timeline

Year 1

Initial six months in Loughborough gaining familiarity with the School, the group and the research methodologies developed at the Loughborough end. Second six months working closely with Landmark, gaining a comprehension of data, machine learning techniques and modelling codes and cloud computing tools.

Year 2

Modelling experiments, generating terrain statistics as a function of time and determining how these vary, as well as how the surfaces respond to key external drivers (uplift rate etc). Processing of 3D seismic data commences. Development of machine learning framework using the stochastic methods.

Year 3

Incorporation of modelling results and seismic analysis into the machine learning framework. Analysis of the efficacy of the data classification that results as well as extracting scientific knowledge from these results.

Further reading

Keylock, C.J. 2017. Multifractal surrogate-data generation algorithm that preserves pointwise Hölder regularity structure, with initial applications to turbulence, Physical Review E 95, 032123, https://doi.org/10.1103/PhysRevE.95.032123.

Keylock, C.J. 2019. Hypothesis testing for nonlinear phenomena in the geosciences using synthetic, surrogate data, Earth and Space Science 6, doi: 10.1029/2018EA000435

Keylock, C.J., Singh, A., Passalacqua, P., Foufoula-Georgiou, E. (2020). Hölder‐Conditioned Hypsometry: A Refinement to a Classical Approach for the Characterization of Topography, Water Resources Research, 56, doi: 10.1029/2019WR025412

COVID-19

This project will employ existing datasets or modelling results generated during the project. As such, it is robust to the impact of COVID-19 of working practices. While the student would benefit greatly from time spent in Abingdon, working with the team at Landmark, it is perfectly possible given the cloud computing capability, to spin-up a virtual machine for the student on which they can undertake any simulation and post-processing, thereby dealing with any potential issues regarding data confidentiality and transfer. Interactions with the supervisory team can all take place via Teams.