A Comparison of Supervised Machine Learning Algorithms to Predict Porosity in Carbonate Reservoirs From Well Logs: A Case Study of Quissamã Formation Carbonate in the Campos Basin, Brazil

Abstract

Porosity estimation is a critical step in characterizing carbonate reservoirs, particularly given their complex pore systems and heterogeneity. This study proposes a data-driven approach to predict porosity in the Quissamã Formation, a carbonate unit in Brazil's Campos Basin, using supervised machine learning techniques. Routine core analysis data served as the target variable, while conventional well logs were used as input features. Eight machine learning algorithms were evaluated: Ridge Regression (RR), Support Vector Regression (SVR), Decision Tree (DT), Random Forest (RF), Extremely Randomized Trees (ET), K-Nearest Neighbors (KNN), and two Multilayer Perceptrons (MLP-1 and MLP-2). The models were optimized through hyperparameter tuning and validated using cross-validation. A total of 328 samples were used, of which 210 (from 3 wells) were allocated for training and validation and 118 (from 5 wells) reserved for blind-well testing, ensuring an independent evaluation of model generalization. Their performance was compared against four traditional porosity estimation methods based on individual and combined well logs. Overall, the machine learning models achieved higher accuracy, with RR demonstrating the most consistent results across wells. This work highlights the potential influence of data representativeness and tuning strategy on prediction performance. Although focused on the Quissamã Formation, the methodology is scalable and adaptable to other reservoirs. Moreover, machine learning models offer practical advantages, requiring no prior petrophysical knowledge and enabling real-time application during well logging, thereby facilitating timely and informed reservoir evaluations.