Introducing AutoML and the noisy-average probabilistic independence of causal influence (PICI) model for the prediction of ornamental stone quality

Abstract

The application of machine learning techniques to analyze large datasets enables mining companies to identify risks and allocate resources effectively, maintaining their competitive advantage. Specifically, software designed to predict the influence of geological factors on ornamental stone is crucial for cost reduction and resource optimization. This study explores an AI-based method to assess roofing slate quality using a noisy-average probabilistic independence of causal influence (PICI) model within an Automated Machine Learning (AutoML) framework. The PICI model was initially introduced and enhanced by incorporating combination functions to reduce model complexity, providing significant advantages in the subsequent phase of Bayesian AutoML inference. This complexity reduction phase was complemented by a t-SNE analysis to visualize data clusters and patterns. The study focuses on the geological setting of the Rodazais Formation in León, northwest Spain, a region renowned for its slate production. Data from 3379 slate sections across 16 boreholes, characterized by 11 different risk factors, were analyzed to determine slate quality. Results indicated that the most influential factors affecting roofing slate quality in the studied deposit were crenulation schistosity, kink-bands, microfractures, and Rock Quality Designation (RQD). Moreover, automated updating of model variables through Bayesian inference and the integration of expert knowledge improved the interpretability of slate quality index predictions and decision-making.