MACHINE LEARNING FRAMEWORK FOR PREDICTING THERMAL PERFORMANCE OF BUILDING ENVELOPES IN EARLY-STAGE ARCHITECTURAL DESIGN

Enhancing thermal performance in architectural design is essential for achieving energyefficient and comfortable built environments. A key determinant of this performance is the building envelope, which encompasses both opaque elements, such as walls and roofs, and transparent components, including windows. These elements influence indoor comfort by controlling heat transfer, air infiltration, and moisture movement, critical factors in architectural engineering. Accurate assessment of how envelope components, especially external walls, affect energy demand requires quantifying heat transfer based on environmental variables (e.g., temperature and humidity) and material properties (e.g., composition, thickness, density). However, physical experimental methods are resource-intensive, and existing simulation tools often require complex inputs, limiting their usability in the early phases of architectural design. This study presents a machine learning (ML) framework designed to support architects and engineers during the early stages of design by enabling the rapid evaluation of wall assemblies. The framework consists of two phases: (1) long-term experimental data collection capturing environmental and material attributes, and (2) training ML models to predict key performance indicators such as heat flux. After preprocessing of datasets (e.g., normalization, handling missing data), various ML algorithms, including random forests and deep neural networks, are benchmarked using root mean squared error (RMSE) as the primary performance metric. The proposed framework offers a fast, scalable, and cost-effective method for integrating data-driven thermal performance analysis into architectural workflows, facilitating better-informed design decisions and improved environmental outcomes.