Abstract
Sustainable mine tailings management has become a worldwide priority given increasing critical raw materials (CRMs) demand and growing environmental concerns. While these anthropogenic deposits are often enriched with useful metals, they may also contain hazardous substances and thus provide both opportunities for resource recovery and environmental risk. In this work a hybrid geostatistical-deep learning framework was established to model geochemical distribution in old tailings. This study integrates ordinary kriging (OK) with a one-dimensional convolutional neural network and a bidirectional long short-term memory model (1D CNN and BiLSTM). The hybrid relies exclusively on features derived from the OK spatial covariance structure, computed from covariance matrices over the sampled locations, to inform the deep model and enhance prediction accuracy. The framework, applied to a historical tailings site, significantly outperformed traditional geostatistical methods as it can provide high-resolution predictions across all points of interest, while accounting for spatial heterogeneity. These results highlight the applicability of this strategy in sustainable resource recovery and environmental remediation, in accordance with circular economy concepts.