Congratulations to Athanasios A. Nathanail with passing his PhD viva with “flying colours” examined by Prof Dorrik Stow and Prof Mikhail Kanevski.
The thesis offers a comprehensive insight into how AI can aid capture deposition interpretational uncertainty from outcrop patterns to make inferences about the conceptual model uncertainty based on the limited in core data and the heritage geological domain knowledge encapsulated in sedimentary literature and reports.
Athos’s work is a fine example of how AI can aid boost efficiency in scaling up routine geological interpretation tasks, quantitative evaluation of interpretational uncertainty and generate added value from consolidate domain knowledge from heritage records.
This thesis demonstrates a novel approach to dealing with interpretational uncertainty by developing an AI system able to learn valuable geological information from surface data (outcrop images), link this knowledge to the fragmented data of the subsurface (core data), and finally, interpret the depositional environment. On the other hand, this AI system can provide a broader range of possible concepts and ideas for the same data because the model has learned from a larger pool of data and data combinations. This overcomes the risk of developing only a limited number of concepts by an expert geologist, with one concept being more prominent than the others.
Specifically, the thesis showcases a Supervised AI system that uses a novel combination of Computer Vision, Natural Language Processing (NLP), and Neural Networks to observe rocks, extract geological knowledge from a corpus of geological data, and embed this knowledge into a custom Neural Network model that combines all the information as a human geologist would into comprehensive interpretations. This approach represents the first attempt to identify the geological depositional environment using only two-dimensional images of sedimentary rocks to generate multiple interpretations, ranked according to the probability of each scenario, to capture the uncertainty. The goal was to create a system that learns from outcrops to apply this knowledge to new outcrop locations and extend this knowledge to other valuable geologic data, including core samples from the subsurface where interpretation is more challenging. Computer Vision algorithms were trained to analyse and segment images of outcrops and automatically learn, identify, and extract features such as rock textures and classify different types of sedimentary structures and lithology types. However, knowledge incorporation into computer vision and machine learning models has been challenging due to the multiple forms of knowledge representation. To make this feasible, NLP is used to elicit expert knowledge from the corpus of geological publications. By creating a customized Neural Network that utilizes the results of both Computer Vision and Natural Language Processing networks, it is possible to generate several different interpretations to predict the likelihood of an outcrop being formed by various depositional environments. The cumulative effect of observing multiple outcrops is to expand the spectrum of scenarios based on geological evidence and fully explore ideas available for interpretation. While the quality of ML-generated output may not surpass that of a human, it can be a valuable tool in providing an indicative overview of geological features.
Athos have shared the finding of his work at various conferences – IAMG, EAGE, and GeoDataScience and UQ knowledge sharing days.