The problem of accurate path loss prediction is still one of the key issues of 5G millimeter-wave (mmWave) implementation, where nonlinear propagation processes have complicated the precision of standardized empirical models. This paper introduces a machine learning system that combines the geometric, temporal, and environmental aspects systematically with the aim of providing a prediction fidelity never before seen. Evaluated on 20,000 propagation samples across 28 GHz and 39 GHz bands in Urban Microcell and Indoor Hotspot scenarios, LightGBM achieves RMSE = 4.21 dB [95% CI: 3.86-4.56] and R² = 0.919 [95% CI: 0.912-0.926] on held out test data representing a statistically significant 29% reduction in RMSE over 3GPP TR 38.901 baselines and 4-10% improvement over alternative ensemble methods (Random Forest, Histogram-based Gradient Boosting, XGBoost). SHAP analysis verifies physical interpretability with the second most important predictor a factor traditionally overweighted in empirical models being RMS delay spread. In practice, the framework achieves 75-percent accurate predictions with the ±3 dB planning margin with the 0.12 ⁻ 1 ms/sample computational cost that is appropriate to real time network planning. The paper makes machine learning a credible upgrade of commercial radio planning systems, allowing covering estimates that are more accurate and lower link margins to implement next generation 5G/6G applications.