The proposed methodology offers several advantages over conventional mineral prospectivity mapping techniques. By using an unsupervised framework, it avoids the need for explicit negative labeling, which can be problematic in underexplored regions where “nondeposit” areas may simply reflect limited discovery. This reduces the risk of introducing false negatives and allows the model to remain sensitive to potentially mineralized zones. The use of a Convolutional Variational Autoencoder also enables the model to learn nonlinear relationships among multiple geoscientific variables. In addition, the separation between western and eastern sectors provides an independent test of spatial generalization across geologically distinct areas. 8.2 Limitations Despite its strengths, the proposed methodology has several limitations that should be acknowledged. A primary limitation is the strong dependence on data quality, consistency, and spatial coverage. Although standard preprocessing steps were applied, residual noise and inconsistencies are likely to remain. Since the model learns directly from the input data, these factors may propagate into the latent representations and influence the resulting prospectivity patterns. Geochemical data were excluded because of their sparse and irregular distribution, which limited their suitability for this framework. The model also used a three-dimensional latent space, which supports interpretation but may restrict representational capacity. In addition, the similaritybased formulation relies on a single reference signature derived from known deposits, which may not capture the full variability of porphyry systems. The training curves showed stable convergence, although the reconstruction term dominated the optimization relative to the KL divergence. This suggests that the latent representation could potentially be further refined. 8.3 Future Work Future work could extend the framework to other deposit types and explore multiple latent signatures within the same model. Additional improvements may include higher-dimensional latent spaces, alternative regularization strategies such as KL annealing, and uncertainty quantification to support more informed exploration decisions. 9. CONCLUSIONS This study demonstrates the application of an unsupervised Convolutional Variational Autoencoder (VAE-CNN) for porphyry mineral prospectivity mapping. By learning patterns directly from multi-source geoscientific data, the approach avoids explicit negative labeling and reduces bias in underexplored regions. The latent-space representation of porphyry systems enabled identification of 49 out of 60 withheld deposits (82%) and delineation of 30 new highsimilarity zones (≥0.90), highlighting areas with geological characteristics comparable to known deposits. Overall, the results indicate that unsupervised representation learning can be applied to porphyry mineral systems and may complement existing methodologies in mineral exploration.
RkJQdWJsaXNoZXIy MTM0Mzk2