Quantitative assessment of interpretational uncertainty in geological mapping with machine learning

点击翻译按钮获取中文摘要

This project aims to tackle the fundamental problem of how to adequately capture and preserve geological uncertainty in reservoirs modelling workflows. Traditional reservoir characterization workflows integrate interpretations from data of different nature - seismic, wireline, core... These interpretations are done by relevant domain experts, who are often separated into siloes, focused only on their aspect of the data, and furthermore can be subject to the experiential bias. When data is finally combined together to create a reservoir model, it is typically done by a single person who must create a coherent representation of the reservoir and its associated uncertainties. Typically, the resulting models are often "best" fits to each bit of data and do not adequately quantify the uncertainties subject to multiple possible interpretations given the sparse nature of the data. Instead we should seek to generate models that combine data in different possible ways to capture the fullest representation of the uncertainty and preserve the knowledge of uncertainties associated with each input data type and its interpretation. The project will develop a way to elicit a wide range of geological concepts as models directly from the data by discovering a variety of data combinations (data types and how they are used/fused together) through using machine learning (ML), while preserving geological knowledge. This will enhance the estimation of uncertainty in our reservoir modelling workflows based on finding a range of unique data combinations/fusions that are coherent (geologically realistic) and unbiased. Achieving this will make a step change enhancement in subsurface uncertainty modelling practice, by identifying a wider possible spread of geological scenarios and providing a quantitative way in assessing their probability. This process will also be significantly faster than a manual approach to developing different ways to use the data to build models. A key element of the project will be to embed geoscience understanding into data driven workflows with machine learning (ML). The challenge is that recent advances in computer science need to be adapted for best implementation in geoscience to capture the context understanding and thinking as it is performed by domain experts. Rigorous ML approach will ensure there is no preferential bias in the way multiple interpretations are elicited from the data. Proper application of ML is able to quantify the impact of the tendencies hidden in the data and how their combinations define possible sedimentological settings. Geologically consistent features need to be derived from the relevant data with account for the associated uncertainty to enhance performance of ML prediction and retain geological realism in ML predictions. The latter requires introduction of solid geoscience understanding into ML prediction model design. Embedded geoscience knowledge will ensure geological consistency of the multiple elicited interpretation leads. The outcome of the PhD project will be a more robust handling of geological uncertainty through novel workflows that use modern machine learning techniques with embedded geoscience understanding. In particular, this will involve adapting Artificial Intelligence (AI) methods, such as deep learning and semi-supervised learning, feature selection, to collate interpretable patterns with the geological context to ensure depositional consistency and geological sense of the predictive models. The feasibility of the novel scientific approach will be justified with a real field study using a modern dataset from the Sea Lion Field, which features modern 3D seismic, as well as log and core data of a comprehensive nature. Once established, this method could be applied to any project where modelling of the subsurface is required.