<p>Slowness is the reciprocal of velocity and known as a standard parameter recorded in sonic logs. Knowledge of compressional (P-wave) slowness (DTC) is vital for structural, geomechanical, and petrophysical analyses of subsurface formations. This study evaluates five machine learning (ML) models, Linear Regression (LR), Decision Tree (DecTr), Random Forest (RF), K‐Nearest Neighbors (KNN), and Support Vector Machine (SVM), to predict DTC in two fractured carbonate reservoirs with differing fracture intensities, referred to as Formation-A (more fractured) and Formation-B (less fractured). A dataset encompassing four wells in each formation was divided into ten incremental sets of petrophysical logs (e.g., GR, RHOB, MSFL, LLS, LLD, RT, PEF, CALD, and CALM). Model accuracy was measured using R² and RMSE across both training and test phases under 1-, 2-, 3-, and 4-Well(s) scenarios. In the training phase, RF consistently attained the highest R² values, up to 0.94 (RMSE≈ 0.26) in Formation-A and 0.92 (RMSE≈0.27) in Formation-B, followed closely by KNN and SVM. In contrast, LR and DecTr showed poor performance in both test and training phases for both formations. Formation-B showed a more stable test performance, often yielding R² of 0.60–0.75 for advanced models. Although data augmentation through more wells and logs typically improved training scores, the best test metrics did not always coincide with the largest feature set (Set #10). Intermediate sets (e.g., Set #5 or #8) sometimes produced stronger generalization, underscoring a trade‐off between model complexity and overfitting risk. RHOB, GR and MSFL emerged as consistently pivotal logs, although deeper resistivity and caliper data also enhanced accuracy under certain conditions. Future research should consider hybrid or ensemble ML methods, data augmentation for underrepresented depth intervals, and the incorporation of seismic attributes to refine sonic predictions in fractured carbonates.</p>