A three-step strategy for integrating contextual and geometric elements in window design optimization through the use of hybrid machine learning

This three-stage model optimizes context-based window design for daylighting, ventilation, energy efficiency, and facade performance. Using machine learning (ML) and architectural reasoning, the suggested method predicts window existence, width, and window-to-wall ratio (WWR) based on geometric inputs such as land size, setbacks, roof proportions, wall offsets, and surrounding conditions. To enhance the model's performance and interpretability, a new Slime Mould Algorithm (SMA) was applied to improve the four ML models of Random Forest (RF), Logistic Regression (LR), Extreme Gradient Boosting (XGB), and Gradient Boosting (GB). Results show that for Stage 1 prediction of window existence, RF generated the best accuracy of 97.99% and F1-score of 0.9685, and had good classification performance for all wall types. In Stage 2, estimation of window width, GB generated the best MAE of 0.1604, R² of 0.8420, and accurately modeled proportional correlations between openings and wall geometry. In Stage 3, the WWR prediction, XGB outperformed regression accuracy with an MAE of 0.0082, R² of 0.9017, tightly correlating contextual density and block layout with facade ratios. SHAP analysis showed roof area ratio affects window presence, while adjacency and density govern WWR. The SMA framework integrates architectural expertise and predictive modeling for sustainable facade design.