<p>The rising demand for high-quality machined parts in precision manufacturing, along with the complexity of machining processes, highlights the need for real-time monitoring to ensure product quality. This study presents a physics-informed methodology for predicting surface roughness parameters in CNC milling, combining tailored feature extraction, machine learning, and interpretability. Force signals were acquired with a piezoelectric dynamometer, while surface roughness was measured online using a chromatic confocal sensor. A lobe-based feature extraction strategy, inspired by chip formation mechanics, was developed to capture localized process information. This was applied to both up- and down-milling with wiper inserts, which are widely used in industry but rarely investigated in surface quality prediction. The engineered dataset was evaluated using a two-step machine learning framework: first classifying milling mode (up vs. down), then predicting roughness parameters. Random Forests and Support Vector Machines were selected for their balance between accuracy and interpretability. Results show that the extracted features encode sufficient physics to support both classification and regression tasks. Predictions were obtained for multiple parameters, including Ra, Rp, Rsk, and Rku, with comparative analysis presented. Beyond accuracy, the approach emphasizes on physics-informed and interpretable modeling, where statistical findings are consistently linked to machining mechanics, and Explainable AI (XAI) ensures model transparency and suitability for online deployment. This methodology bridges machining science and AI, aligning with the objectives of zero-defect manufacturing (ZDM).</p>