The global transition toward low-carbon energy systems has significantly increased the demand for critical metals, particularly copper, which plays a fundamental role in electrification, renewable energy infrastructure, and energy storage technologies. According to the International Energy Agency, demand for copper and other energy transition minerals is expected to grow substantially in the coming decades, placing increasing pressure on the mineral exploration sector to identify new economically viable deposits (IEA, 2021). Despite this growing demand, mineral exploration remains a high-risk and capitalintensive process, especially in remote and underexplored regions such as the Yukon. Traditional exploration methods, including geological mapping, geochemical sampling, and geophysical surveys, have historically been effective in discovering mineral systems. However, these approaches are often constrained by high operational costs, logistical challenges, and limited spatial coverage, particularly in areas with difficult access or sparse infrastructure (Singer & Kouda, 1999). In recent years, data-driven approaches have been introduced to improve exploration efficiency. In particular, supervised machine learning techniques have been widely applied to mineral prospectivity mapping, where models are trained to distinguish between known deposits and non-deposit locations (Xiong et al., 2018; Sun et al., 2019). While these methods have demonstrated promising results, they rely heavily on the availability and quality of labeled datasets. This dependence introduces a fundamental limitation in underexplored regions, where the absence of known deposits does not necessarily imply the absence of mineralization. Consequently, areas labeled as “non-deposit” may in fact represent undiscovered targets, leading to the introduction of false negatives and bias in model training (Cracknell & Reading, 2014; He et al., 2024) In this context, unsupervised learning approaches offer an alternative framework that does not rely on explicit labeling. Among these, Variational Autoencoders (VAEs), introduced by Diederik P. Kingma and Max Welling (Kingma & Welling, 2014), provide a probabilistic framework for learning latent representations of complex, high-dimensional data without requiring labeled outputs.
RkJQdWJsaXNoZXIy MTM0Mzk2