ABSTRACT
Uniaxial Compressive Strength (UCS) is an essential and critical mechanical parameter for design and stability assessment in rock engineering. In practice, accurate measurement of the UCS of rocks is obtained through laboratory-based direct testing, however, these are expensive, time consuming, and destructive. To overcome these shortcomings, numerous researchers have employed indirect methods to generate UCS-predictive models using soft and hard computing methods. Although, these indirect models remain limited due to insufficient availability of data restricting their applicability to different sites and lithologies. Therefore, alternative methods to overcome this shortage of data are necessary and of a high interest. In this study, an Artificial Neural Network (ANN), a Decision Tree (DT), and Multi Linear Regression (MLR) model are used to develop UCS predictive models. To this end, an experimental dataset of nineteen UCS, point load tests and bulk density measurements on sandstone was built and three statistical metrics were used to evaluate each model performance. Subsequently, Transfer Learning (TL), a technique which allows smaller datasets to be used, is applied to ANN to test its effectiveness. The results showed that TL achieved superior predictive performance on unseen data compared to the ANN, DT and MLR models and can be effectively applied to enhance the performance of a model built with a small dataset.
