Anisotropic Marble Quality Estimation Using Orientation-Aware Neural Networks

Reliable estimation of marble quality is a key prerequisite for effective quarry design, production scheduling, and economic assessment. In ornamental stone deposits, quality attributes such as veining intensity, colour variation, and textural patterns exhibit pronounced directional dependence, meaning that the apparent quality of a block or sample is strongly influenced by cutting orientation. Drillhole data are acquired along multiple trajectories in three-dimensional space, and the resulting observations implicitly encode this directional dependence, which is commonly neglected in conventional spatial estimation workflows. To address this limitation, this study introduces an orientation-aware machine learning approach for the prediction of marble product proportions based on neural networks. The proposed framework integrates spatial coordinates with explicit orientation descriptors, represented through trigonometric encodings of azimuth and dip, alongside multi-scale Random Fourier Feature transformations that enable the network to capture complex, anisotropic spatial patterns. The model is trained on composite drillhole data and subsequently deployed to generate predictions across a three-dimensional block model. Permutation-based interpretability analysis indicates that orientation-related inputs contribute substantially to predictive performance, in many cases exceeding the importance of spatial coordinates alone. These results demonstrate that explicit incorporation of orientation information enables the model to learn anisotropic quality behaviour that would otherwise remain unresolved, leading to more accurate and operationally meaningful predictions. The methodology is illustrated through a case study from a marble quarry in northeastern Greece, highlighting its practical relevance for resource evaluation and production planning.