Development of Supervised Machine Learning Models for the Prediction of Well-Logs & Application on Wells at São Francisco and Santos Basins, Brazil

Porosity assessment is essential for reservoir characterization. While laboratory measurements and conventional well-logs have traditionally been used to estimate porosity, they may not always provide accurate results in carbonate reservoirs, due to their extremely heterogeneous pore system. This is where nuclear magnetic resonance (NMR) tools offer a valuable solution. NMR logging technology allows for more accurate quantification of different porosity types, including total, effective, and free-fluid porosity. However, acquiring NMR logs can be costly and challenging due to factors such as the activation of wireline equipment, signal-to-noise ratio, environmental factors, and the properties of the formation fluid. To overcome these challenges and provide an alternative approach, we are interested in developing predictive models, using conventional well-logs as input data. The objective of this research is to develop a Python code, from scratch, that implements supervised machine learning (ML) algorithms, specifically Random Forest (RF) and Gradient Boosting (GB), to build ML models for accurately predicting both NMR logs and various types of conventional well-logs. The code is available as an open source on GitHub under the repository named “Well-Logs_Predictive_Models” [https://github.com/VittoDePe98/Well-Logs_Predictive_Models.git]. This open accessibility encourages wider usage and collaboration among researchers. Two separate case studies are conducted to evaluate the functionality and effectiveness of the code. In both studies, the ML models are trained on a first well (training well) and tested on a second well (test well). The first case study focuses on the São Francisco onshore Brazilian basin. It serves as a preliminary exercise for model development and data familiarization. The goal is to predict two conventional well-logs: the calculated effective porosity and the measured compressional wave slowness logs. The second case study centers around the Santos offshore Brazilian basin, particularly the deep-water pre-salt carbonate reservoir area of the Itapu Oil Field. The attention of this research is primarily directed toward this second case study. The target is to predict high technological well-logs, including NMR total, effective, and free fluid porosity logs. The results of the research demonstrate that both models are consistent and reliable, exhibiting low regression errors (MSE, RMSE, MAE) and high accuracy (R2) values, in predicting the calculated effective porosity, for both training and test wells. However, when it comes to predicting the measured compressional wave slowness log, the models exhibit limitations and show poor performance on the test well. The limitations become even more apparent when predicting NMR porosity logs on the evaluation well. The models exhibit significantly reduced performance, yielding negative accuracy values.