Utilização de Machine Learning supervisionado para a predição de litofácies no pré-sal da Bacia de Santos, formação Barra Velha, Campo de Tupi

Abstract

Lithofacies classification is crucial for the exploration and development of oil and gas fields, providing valuable information for petrophysical and seismic analyses. In the case of carbonate lithofacies, the heterogeneity across different scales poses a challenge, particularly because manual interpretation by experts is time-consuming and prone to bias. In Brazil, the discovery of large hydrocarbon fields in the pre-salt reservoir interval has generated significant interest in the origin and classification of complex carbonate rocks. Recently, machine learning methods have emerged as a promising solution for automated lithofacies classification processes. In this study, we present an approach that includes the use of the XGBoost algorithm and a dataset from the Barra Velha Formation in the Tupi field, Santos Basin, Brazil, to perform high-precision automated classification of carbonate lithofacies. The methodology involved data selection, quality control, and preprocessing, followed by data augmentation, hyperparameter tuning for each model, training of the XGBoost algorithm, and model evaluation using important metrics in the stages of training, cross-validation, and testing. Twelve different models were evaluated, with four of them being selected for their application in blind wells. These models achieved F1-scores ranging from 0.63 to 0.72. The results demonstrate the effectiveness of the XGBoost algorithm combined with advanced variable selection, data augmentation process, and hyperparameter tuning. The approach and workflow created with the dataset from the Barra Velha Formation of the Tupi field provided valuable insights for exploring the pre-salt interval in Brazil, as well as for its promising application in classifying lithofacies in different carbonate basins, such as the Campos and Kwanza basins.