Explainable Sustainability Intelligence for Manufacturing Organizations Using a SHAP-Based AI Decision-Support Framework

Sustainability assessments in manufacturing typically report on performance but provide little evidence of which observed indicators influenced a near-term outcome, or which improvement should be prioritized by management. The published sustainable industrial measurement (SIM) model is expanded into an explainable sustainability intelligence framework that links expert-validated indicators and analytic hierarchy process (AHP) weights to one-month-ahead XGBoost prediction, SHAP explanations, and sensitivity-tested improvement prioritization under human review. Three synthetic factories contributed monthly observations during 2020–2024. Fourteen expert-validated SIM indicators were used as predictors. The XGBoost model achieved R2 = 0.9598, RMSE = 0.2248, and MAE = 0.1431 during validation, and R2 = 0.9685, RMSE = 0.1492, and MAE = 0.0791 on the 2024 temporal holdout. SHAP analysis showed that the share of renewable energy (38.74%), the waste recycling rate (24.08%), the CO2 emission intensity (12.83%), and the material efficiency (12.45%) were the main predictors, accounting for 88.10% of the total mean absolute SHAP importance. Explanations of the sustainability profiles were factory-specific and distinguished the transitional, improving, and advanced profiles for Factories A, B, and C, respectively. The framework decouples data-driven predictive reliance from expert-derived strategic importance, preventing feature importance from being an automatic intervention recommendation, and transforming prediction, local attribution, benchmark potential, and circularity relevance into an auditable human decision sequence. The target was a deterministic weighted score of seven environmental indicators. Persistence outperformed XGBoost (test RMSE of 0.1228 versus 0.1492), which identifies a clear role for this proof of concept: making forecasts, explanations, and review priorities traceable before operational validation.