Abstract: The transient-based leak detection (TBLD) has been established as a robust and efficient technology to manage and monitor pipe systems. The TBLD is a non-convex multi-dimensional optimization problem and is usually solved by heuristic optimization algorithms (HOAs). HOAs are sensitive to noise, susceptible to convergence to local optima, and computationally expensive. Motivated by the need for a reliable and practical approach, this study put forward a novel machine learning (ML)-based framework to detect leaks in pipes. At the core of this methodology, an ensemble of CatBoost models was trained on more than 3.8 million data records for classifying the leaky sections and predicting the leak sizes in them. Our results showed that the ML models can detect leaks with 97% accuracy and an F1-score of 0.86. Followingly, the ML models can detect the size of leaks with R2=0.93 and RMSE=0.025 of the pipe area. Supported by quantitative analysis, the ML models could significantly outperform an HOA algorithm in terms of accuracy and computation time. The proposed framework enables practitioners to detect leaks in hydraulic systems reliably and efficiently. By substituting the cumbersome optimizations with an ML-based approach, this study opens a new line of research in the TBLD field.