Machine Learning-Based Depth Selection For Enhancing Wireline Formation Testing Efficiency In Complex Siliciclastic Reservoirs

Paleozoic siliciclastic reservoirs, such as those in the Amazonas and Parnaíba Basins (Northern Brazil), exhibit high vertical and lateral heterogeneity. This complexity poses significant challenges for selecting measurement depths for Wireline Formation Testing (WFT). Poor test point selection often leads to non-representative pressure data, compromising reservoir characterization, increased operational costs and security risks. To mitigate these risks, this study proposes two complementary Machine Learning (ML) approaches that rely exclusively on conventional well logs (GR, PE, RHOB, NPHI, DTC, DTS, shallow and deep resistivity). The objective is to support near real-time decision-making during logging operations. Two models with distinct, yet complementary functions were developed. Model 1 performs a binary classification to predict the effectiveness of a WFT station. Model 2 classifies points into four mobility categories (dry, low, regular, high), prioritizing zones with a higher probability of rapid pressure stabilization. Various algorithms were tested, including K-Nearest Neighbors (KNN), Random Forest, Artificial Neural Networks, XGBoost, and CatBoost. CatBoost demonstrated the best performance for Model 1 (ROC AUC ≈ 0.94, Accuracy ≈ 0.86, MCC ≈ 0.74) and the same for Model 2 (ROC AUC (macro) ≈ 0.87, Accuracy ≈ 0.73, MCC ≈ 0.61) in cross‐validation. Both models were validated using a Group K‐Fold by well (k = 10) to prevent leakage between training and testing wells, followed by blind‐well evaluation, and underwent extensive hyperparameter optimization. The models were successfully applied and validated across two analogous intracratonic basins (Amazonas and Parnaíba),coveringmultiple formationsanddepositional contexts. Blind‐well resultsfurther confirmed robustness and generalization (e.g., binary model with ROC AUC ≈ 0.92 and MCC ≈ 0.57). Operationally, Model 1 significantly reduces the occurrence of unsuccessful pressure tests, while Model 2 guides the selection of points for higher testing efficiency, enabling faster and more assertive field decisions. However, the contrast between the vertical resolution of the logs (around 15 cm) and the test point limitations may reduce the methodology accuracy in highly heterogeneous facies, such as heteroliths, where centimeter‐scale variations affect mud cake formation and the tool sealing. This study demonstrates that applying machine learning directly to conventional well logs significantly optimizes WFT depth selection, leading to reduced costs, non‐productive time,andoperational risks. Its successful application in two distinct Paleozoic siliciclastic basins (Amazonas and Parnaíba), covering multiple formations and depositional contexts, confirms the robustness and generalization capability of the models. This scalability lays the groundwork for future integration with image logs and advanced machine learning techniques, expanding applicability to a broader range of depositional environments. Despite limitations in vertical resolution, the models offer strategic value for field operations, especially in complex siliciclastic reservoirs.The complementarynature ofthe twomodels, one focused on test effectiveness and the other on mobility classification, ensures a more comprehensive and reliable decision support system for wireline formation testing.