Application of Machine Learning Models to Select Quarry Limestone for Heritage Restoration

Abstract

The fortification system of Cartagena de Indias, designated as a UNESCO World Heritage Site in 1984, represents a fundamental element of the city’s cultural heritage and historical legacy. However, significant portions of these structures have undergone severe degradation due to environmental and anthropogenic factors, requiring continuous restoration efforts. Currently, the selection of replacement limestone blocks is conducted without a standardised technical criterion, potentially compromising the structural integrity and breathability of the walls. Stones with high porosity exhibit low mechanical resistance, making them susceptible to failure under static and seismic loads. In contrast, those with low porosity hinder moisture exchange, leading to overpressure and biodeterioration. This study introduces a machine learning-based (ML) approach to predict the Uniaxial Mechanical Strength (UCS) both in dry and wet conditions of quarry limestone intended for restoration, utilising key physical parameters, such as real volume, average wave speed, saturated mass, submerged mass, and dry mass. Seven predictive models were evaluated: Ensemble Methods (EM), Gaussian Process Regression (GPR), Kernel-based Regression (KR), Linear Regression (LR), Neural Networks (NN), Stepwise Linear Regression (SLR), Support Vector Machines (SVM), and Tree-based regression (TR). The novelty lies in applying these models to improve material selection for restoration, ensuring the structural and aesthetic integrity of the restored limestone structures while minimising the need for destructive testing. The findings reveal that SLR has the highest predictive accuracy across both dry and wet conditions, achieving an R² of 87% with an RMSE of 3.42 for dry UCS, and an R² of 85% with an RMSE of 2.44 for wet UCS. Furthermore, Ensemble, Gaussian Process Regression, and Quadratic SVM models demonstrated significant performance, achieving a 5% improvement in R² compared to standard LR models. Finally, an original criterion for limestone selection as a replacement for cultural heritage restoration is proposed as a tool for competent authorities.