f Facies classification using machine learning: lessons from SEG-ML contest

New resources exploration and exploitation sites collect at high resolution and rate multiple sources of data, generating considerable amount of information to process and eventually, interpret. Newly developed Machine Learning algorithms can help overcoming this challenge to gain better insight on data and resources. As the latter is hot topic right now in geoscience, Matt Hall, Editor of Leading Edge’s Geophysical Tutorial launches a contest for facies prediction in 2016 October issue. For this purpose, wireline logs and geological facies data from nine wells in the Hugoton natural gas and helium field of southwest Kansas (Dubois et al. 2007) were made available, with facies in two wells being kept blind to contestants. With contributions from data scientists from all over the world, the best score achieved was a little over 63% for the automated prediction of 9 different facies. Submissions included a dozen different algorithms, but ensemble methods (Gradient Tree Boosting, Random Forest) proving to be the most successful ones. Also, feature engineering (the extraction of more information from the variables) turned out to be the key aspect for successful prediction. In addition, comprehensive interpretation of the contestants results showed that prediction uncertainty was roughly about 4%.