Machine learning approaches using historic materials' data for hygrothermal modelling

The intersection of machine learning and cultural heritage preservation represents a new frontier in conservation science. This thesis highlights the potential of artificial intelligence in safeguarding our monuments. This thesis investigates the application of machine learning to predict hygrothermal properties of historical building materials using data from mercury intrusion porosimetry (MIP) tests. The study focuses on developing models to estimate the capillary water absorption coefficient (Acap) and capillary moisture content (Wcap) of brick samples, utilizing features derived from MIP data. Conducted in partnership with the Royal Institute for Cultural Heritage (KIK-IRPA) in Belgium, this research utilizes a comprehensive dataset of 706 brick samples, tested for both MIP and water capillary absorption rate collected over the period from 2007 to 2021. After data cleaning and preprocessing, random forest regression models were implemented to predict Acap and Wcap. Two sets of input features were compared: one including the coordinates of the pore size distribution peak, and another without these coordinates. Results indicate that the models predict Acap with moderate accuracy (R2 up to 0.665) while predictions for Wcap were less reliable (R2 up to 0.275). Notably, the inclusion of pore size distribution peak coordinates significantly improved model performance for Acap prediction. This research contributes to the field of building conservation by proposing innovative, data-driven and less destructive methodologies to estimate important hygrothermal properties of historical materials. The findings suggest that machine learning models, particularly those incorporating detailed pore structure information, can provide valuable insights for material characterization, preservation strategies and more tailored and informed approaches in architectural heritage conservation.