Abstract A unified unsupervised-learning engine is introduced that requires no labelled data yet discovers structure in large, heterogeneous geoscience datasets. Two representative applications from a wider portfolio are demonstrated: (i) regional clustering and ranking of copper-prospective ground from magnetic, gravity and topographic grids, and (ii) identification of key geologic features—faults, channels and stratigraphic facies—within the publicly available Netherlands offshore F3 three-dimensional seismic volume. The goal is to show that one adaptable technology can support multiple exploration workflows, maximise known-target recall and shrink the area or volume demanding costly follow-up work. Raster inputs are Z-score normalised and patchified into 64 × 64-pixel (or trace-window) vectors, then processed through a three-branch ensemble: (1) a Deep Embedded Clustering autoencoder that learns latent features and soft memberships; (2) a 10 × 10 Self-Organising Map whose code-book vectors are re-clustered with K-Means using the Davies–Bouldin elbow; and (3) a UMAP→HDBSCAN stream that captures variable-density structure. Branch outputs are fused by majority vote, while per-sample entropy provides an uncertainty score. Resulting cluster rasters or SEG-Y masks are geo-referenced for direct overlay with structural lineaments, lithology, well control or interpreted horizons. Reproducible configuration files allow rapid transfer to other commodities or survey areas. Across two copper-bearing provinces in India the system retrieves 95 % of mapped porphyry and volcanogenic massive-sulphide occurrences while trimming the prospective footprint by 30 % relative to a SOM-only baseline. On the F3 seismic dataset it automatically delineates the principal fault network, four dominant facies belts and a previously under-reported channel complex; interpreter cross-checks show a 24 % gain in feature-detection precision and a 19 % reduction in false positives compared with manual trace-gating. Uncertainty maps—calibrated to a Brier score of 0.12 versus 0.21 for the baseline—consistently steer follow-up geophysics and drilling toward the highest-confidence targets. The technology is the first to unify Deep Embedded Clustering, SOM-K-Means and UMAP–HDBSCAN in a single, scalable engine and to validate it quantitatively across both mineral and hydrocarbon domains. Relative to standalone SOM, the ensemble cuts computational time by 42 %, lowers the Davies–Bouldin index by 18 % and boosts cluster stability (bootstrap Adjusted Rand Index +0.15). Integrated uncertainty quantification reduces high-risk acreage by 40 % while preserving 98 % of known positives, translating advanced unsupervised learning into confidence-graded targets for real-world exploration decisions.
Kumar et al. (Mon,) studied this question.