Depositional Facies Prediction Using Artificial Intelligence to Improve Reservoir Characterization in a Mature Field of Nam Con Son Basin, Offshore Vietnam

Abstract Interpretation of geological facies such as lithofacies and depositional facies is a process of classifying high dimensional data into groups. Conventional Machine Learning methods such as random forest, support vector machine, back progradation neutral network have been tested throughout in lithofacies classification (Hall and Hall, 2017). However, poor performances were observed in depositional facies prediction as the later requires integration of more complex data such as depositional facies interpretation in core, wireline, biostratigraphy in the context of depositional units, not sample by sample data points. The limitation of data-driven classifier and the value of geological derived features in facies classifications was discussed in the literature (Halotel et al., 2020). In order to improve the classifier performance, a novel hybrid approach was implemented, which involved an automatic depositional unit breakdown, well log feature extractions, machine learning core-log models and a rule based expert system. In this paper, an Artificial Intelligence (AI) system was employed that integrates machine learning with an expert system to predict depositional facies and tested its reliability versus facies interpretations made from conventional cores in test wells. In the prediction method, data samples were depositional units derived from Gamma Ray value changes that were integrated with core and biostratigraphic data. These were then passed through an expert system containing a set of fuzzy rules to yield final probabilities for each of the 35 depositional facies stored in a library. The AI system helped to produce multiple scenarios with qualify uncertainty, consistent depositional facies classification based on depositional facies observed in core, biostratigraphy and wireline logs. Depositional facies in the aforementioned reservoir were predicted with the AI system for the cored intervals in X1 and X2 using only well logs and biostratigraphic data as input. The predicted facies were in strong agreement with the core interpretations as both methods concluded that the depositional environment was a marine delta. The AI prediction assigned a higher proportion of the X2 succession to the "Marine Delta" depositional facies than it did for the X1 core; the facies difference in the two wells was subtle and not recognized by the core interpretation. Seismic attributes suggested that the more variable facies succession that AI predicted for X1 may have geological significance. The AI tool generated reliable and consistent results and appeared capable of reducing the uncertainties of predicting facies distribution and the subsequent development of conceptual depositional models.