Accurate storm surge forecasting along India's 7500 km coastline remains a significant challenge due to the frequent occurrence of tropical cyclones and the complex dynamics of its coastal regions. While machine learning (ML) offers a computationally efficient alternative to hydrodynamic models, conventional approaches often lack physical interpretability and struggle to predict rare, high-impact events. This study introduces a robust, pan-India physics-guided machine learning (PGML) framework for 6- to 48-h extreme surge forecasting. Using ERA5 reanalysis and IBTrACS cyclone datasets across 229 tide-gauge stations, we benchmarked traditional Gradient Boosting models (LightGBM, XGBoost, CatBoost) against novel PGML variants and stacked ensemble architectures. The Stacked Physics-Guided (Stack-PGML) model consistently achieved superior accuracy, delivering the lowest Root Mean Squared Error (RMSE = 0.042 m at 24 h; R2 = 0.80) under both general and extreme-event conditions. SHAP interpretability analysis confirmed the model's physical realism, identifying sea level pressure as the dominant predictor and validating the inverse barometer effect, addressing the “black box” limitation of standard ML. The results demonstrate that integrating physical constraints and ensemble learning yields a scalable, accurate, and interpretable approach for data-driven storm surge estimation across the Indian coastline.