Machine Learning Supported Risk Revision for Seismic Amplitude Anomalies and Implications for an Exploration Portfolio

Summary The chance of geological success (Pg) is a vital metric across oil & gas industry. Prospects with Direct Hydrocarbon Indicators (DHIs) are reported to have higher success rates ( Roden et.al. 2014 , Rudolph and Goulding, 2017 ), hence optimal DHI assessment and selection of appropriate method is vital for management of exploration portfolio. In the current study, we investigate how additional information on DHIs can affect the revision of Pg. Three models for comparison are considered: Initial Pg, Revised Pg with Bayesian update and Revised Pg with direct application of machine learning algorithm. Binary classification metrics, such as confusion matrix, balanced accuracy along with confidence interval, sensitivity and specificity have been used to select the best predictive model. All three models are compared by analyzing density probability plots and receiver operator characteristic (ROC) curves for Revised Pg. Revised Pg’s with encountered DHI information show higher discriminating ability in comparison to Initial Pg, supporting historically reported higher success rate for DHI-associated prospects. Selection of the revision method, whether it is Bayesian or machine learning update, significantly affect the distribution of Revised Pg for the company’s prospects portfolio. Decision on the appropriate method can be done based on the company’s risk tolerance strategy.