Machine Learning assisted carbonate classification
George Ghon, Geo and data Scientist at Earth Science Analytics
17 January 2022
Despite what is traditionally a time where many of us take an extended summer holiday, we were delighted to welcome attendees to our latest webinar in July: Machine Learning Assisted Carbonate Lithofacies Classification - A Barents Sea Case Study.
George Ghon, a Geo and Data Scientist at Earth Science Analytics presented findings from a recent study of an upper Palaeozoic carbonate factory in the Barents Sea.
The study utilised ML assisted lithofacies classification through novel rock typing approaches to facilitate reservoir characterisation. Those in attendance learned about how ML models can be trained with high quality labels to guide quantitative interpretation from wellbore data.
George outlined the Loppa High, while referencing the recent Alta and Neiden discoveries. For the Alta discovery, George shared information on the reservoir’s characteristics including the eastward tilt of the basement high and palaeozoic sediments, collapsed breccias and conglomerates.
For context, information was also presented on the tectonostratigraphic overview of the Loppa High, including its upper Carboniferous and lower Permian carbonate factories transgressing from tropical to cool water environments during those times. The regional lithology was also outlined and shown in cross plots with well log data to monitor clustering trends.
Those on the webinar then learned about the ML workflow which centres initially around core labelling and well log features. From there, George noted the importance of undertaking a pre-ML study, before introducing a ML classification algorithm. George presented his conclusion and outlook on the findings which centred around four key points.
Firstly, upper Palaeozoic carbonate reefs are heavily overprinted, with reservoir properties controlled by deformation, erosion, and diagenesis. In those environments, contextualised labelling strategies can significantly improve classification.
We have noted in previous blogs that even with advancements in data labelling, insight from subject-matter experts via labelling is still required in the quest to deliver accurate models. This point was reiterated during the presentation of the Barents Sea activity.
This led to a discussion and the hypothesis that so-called neural nets, or what would be more commonly known as MLP classifiers, have the potential to predictively quantify small scale reservoir heterogeneities better than linear classifiers.
He closed by presenting the argument that the advantage of machine learning over traditional approaches becomes increasingly important with bigger data applications.
We have many more webinars to come this year, so please follow our LinkedIn page for information and to sign up.
Alternatively, if you would like more information on any of the topics noted above, please don’t hesitate to get in touch with us via our contacts page.