<p>Building material stock studies are essential for advancing the circular economy in construction. However, existing models often lack both accuracy and scalability. While machine learning has demonstrated significant potential to enhance predictive accuracy, its adoption has been hindered by a shortage of high-quality training data. In this study, we introduce a novel methodology leveraging a large language model to extract previously untapped building material data from building energy performance certificates with a focus on exterior walls. This approach enabled us to create a dataset of over 20,000 buildings—significantly larger than those used in previous studies. Leveraging this dataset, we developed a machine learning model to predict material composition based on building characteristics such as construction year, use, and location. Furthermore, we integrated knowledge of construction history to estimate the material stock of walls in terms of volume, mass, and associated CO₂ emissions for each building in the dataset. Our analysis revealed significant regional variations in material use patterns, emphasizing the critical role of location—a parameter often overlooked in existing building material stock models. These findings provide valuable insights for improving building stock modeling and highlight the importance of regionally tailored policies in advancing the circular economy in the construction sector.</p>