Integración de kriging y aprendizaje híbrido (Autoencoders k-Prototype) para la definición y mapeo 2D de dominios geoestadísticos de Manganeso (Mn) y Fósforo (P)

Abstract

Abstract: This research developed and validated a hybrid methodology for the objective definition of geostatistical domains for Manganese (Mn) and Phosphorus (P), integrating dimensionality reduction through an Autoencoder, mixed-data clustering with k-Prototype, and validation through ordinary kriging. From a multielement geochemical dataset, latent factors were generated to summarize the main variation patterns of 35 predictor variables, which were then combined with lithology to define geochemical-lithological domains. The performance of these domains was compared against a global model without domains and a conventional scheme based only on lithology. The results show that the hybrid approach produced a substantial and statistically significant improvement for Phosphorus (P), reducing validation errors and increasing predictive capacity. In contrast, Manganese (Mn) showed a marginal response, with performance comparable to the global model and no statistically robust improvement. Overall, the study demonstrates that domains generated through hybrid learning can improve geostatistical estimation, but their effectiveness depends on the target variable and the geological controls that condition its spatial continuity.