Ground-penetrating radar has been used in pavement inspection for layer evaluation and anomaly detection. However, the inversion of radargram responses into quantitative crack geometry remains unresolved, particularly for narrow discontinuities expressed as localized hyperbolic signatures. This study presents a descriptor-based machine learning framework for estimating pavement crack width w and crack depth d from synthetic GPR data generated under controlled numerical conditions. A parametric dataset was generated in gprMax for top-down and bottom-up cracking configurations, varying crack width from 2 to 20 mm, crack depth from 2 to 10 cm, bituminous layer thickness from 5 to 20 cm, and the geometric parameter h_g from 10 to 50 cm. The simulated B-scans were processed in MATLAB to identify crack-related hyperbolic responses and to extract physically interpretable descriptors linked to temporal and amplitude behavior, including upper-event and lower-event two-way travel time, reflected amplitude, temporal separation, amplitude difference, and amplitude ratio. The resulting feature matrix was exported to Python, where Linear Regression, Support Vector Regression, and XGBoost were trained to estimate w and d. Model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). Descriptor availability analysis showed that 84.4% of the simulations yielded a complete upper–lower event pair, whereas 15.6% did not present a detectable lower event under the adopted tracking criteria, with a higher occurrence in bottom-up cases. The best width estimation was obtained with Linear Regression, with R² = 0.723, RMSE = 2.455 mm, and MAE = 1.410 mm. The best depth estimation was obtained with XGBoost, yielding R² = 0.535, RMSE = 1.339 cm, and MAE = 0.654 cm. Although developed on synthetic data, the proposed workflow provides a physically interpretable basis for future validation with laboratory and field GPR measurements.