An explainable machine learning framework for multi-objective carbon reduction targeting material operational seasonal emissions in building retrofits.

As the global initiative for carbon neutrality in the construction sector accelerates, the low-carbon retrofitting of existing buildings is emerging as a critical pathway to combat climate change. This paper proposes a systematic framework that integrates explainable machine learning with multi-objective optimization to support the sophisticated optimization of carbon emissions in renovation projects. The framework is centered on three core. Material Carbon Emission Intensity (MCEI), Operational Carbon Emission Intensity (OCEI), and Seasonal Carbon Emission Balance (SCEB). Leveraging high-resolution carbon emission simulation data, predictive models were developed using six machine learning algorithms, among which CatBoost demonstrated superior performance. Subsequently, SHAP values were employed to identify key design variables influencing carbon emissions, such as FLH, WWR1, NOF, and WWR2, thereby providing an evidence-based foundation for strategic decision-making. The framework's utility was validated through a case study of a three-story industrial building retrofit, where the NSGA-II algorithm was applied for multi-objective optimization. This process yielded four distinct sets of feasible solutions. The most balanced solution achieved a 71.06% reduction in MCEI, a 37.20% reduction in OCEI, and a 24.75% improvement in SCEB compared to the baseline scenario. This study culminates in a series of recommended low-carbon strategies, including material reuse, promotion of low-carbon materials, optimization of partition walls, enhancement of the thermal performance of the building envelope, and improvements in atrium design. In conclusion, this research provides a systematic, scalable, and replicable technical pathway for the low-carbon retrofitting of buildings, holding significant practical value for achieving carbon neutrality goals.