Summary Knowing the location of sweet spots benefits the horizontal well drilling and the selection of perforation clusters. Generally, geoscientists determine sweet spots from the well-logging interpretation. In this paper, a group of prevalent classifiers [extreme gradient boosting (XGBoost), unbiased boosting with categorical features (CatBoost), and light gradient boosting machine (LightGBM)] based on gradient-boosting decision trees (GBDTs) are introduced to automatically determine sweet spots based on well-log data sets. Compared with linear support vector machines (SVMs), these robust algorithms can deal with comparative scales of features and learn nonlinear decision boundaries. Moreover, they are less influenced by the presence of outliers. Another prevailing approach, named generative adversarial networks (GANs), is implemented to augment the training data set by using a small number of training samples. An extensive application has been built for the field cases in a certain oilfield. We randomly select 73 horizontal wells for training, and 13 features are chosen from well-log data sets. Compared with conventional SVMs, the agreement rates of interpretation by XGBoost and CatBoost are significantly improved. Without special preprocessing of the input data sets and conditional tabular GAN (CTGAN) model fine tuning, the fake data set could still bring a relatively low agreement rate for all detections. Finally, we propose an ensemble-learning framework concatenating multilevels of classifiers and improve agreement rate. In this paper, we illustrate a new tool for categorizing the reservoir quality by using GBDTs and ensemble models, which further helps search and identify sweet spots automatically. This tool enables us to integrate experts’ knowledge to the developed model, identify logging curves more efficiently, and cover more sweet spots during the drilling and completion treatment, which immensely decrease the cost of log interpretation.