<p>Lightweight ultra-high-performance fiber-reinforced cementitious composites (L-UHPFRC) have attracted increasing attention in modern construction due to their superior mechanical performance and reduced self-weight, making them highly suitable for long-span bridges and high-rise structures. However, accurately predicting their compressive strength (CS) remains challenging because of the complex nonlinear interactions among constituent materials. This study proposes an interpretable machine learning framework integrating a hybrid Random Forest–Whale Optimization Algorithm (RF–WOA) model for predicting the CS of L-UHPFRC. The proposed RF–WOA model demonstrated high predictive accuracy, achieving R² = 0.934, RMSE = 6.816, MAE = 4.698, and MAPE = 3.410% for the training dataset, while also exhibiting stable generalization performance with prediction-to-experiment ratios of (1.01, 0.05) and (0.99, 0.08) for the training and testing datasets, respectively. Furthermore, a novel SHAP-based explicit formulation was developed to capture the nonlinear relationships between input variables and CS while improving model transparency and engineering interpretability. Partial Dependence Plot (PDP) analyses further confirmed that the learned relationships are physically consistent with the underlying behavior of L-UHPFRC. The results demonstrate that the proposed framework provides a reliable, accurate, and interpretable tool for compressive strength prediction and mix design optimization of advanced lightweight cementitious composites, thereby supporting their broader practical implementation in sustainable and high-performance structural applications.</p>