Enhancing porphyry copper prospectivity mapping: A deep autoencoder-based approach to identify non-deposit points in varzaghan region, NW Iran

Abstract

Supervised machine learning algorithms have shown enormous potential to predict mineral prospectivities and to identify mineral exploration targets within study areas. However, accurately selecting non-deposit points remains a critical challenge, as improper selection can mitigate the prediction rate and introduce systematic bias. This study describes the idea of leveraging and comparing deep autoencoder (DAE) network (as a first experiment) with expert knowledge (as a second experiment) to tackle the problem of non-deposit selection in predictive modeling of mineral prospectivity. For this, according to the conceptual model of porphyry copper deposits evidence layers of fault density, multi-element geochemical signatures, proximity to phyllic and argillic alterations, and proximity to intrusive rocks, were first generated to represent ore-forming subsystems. Within the first experiment, a DAE technique was used to integrate multiple exploration criteria whereby non-deposit locations within the recognized non-prospective regions were determined. Within the second experiment, expert opinions were set as criteria to define non-deposit locations. Both sets of non-deposit points were fed into a random forest (RF) algorithm, generating two prospectivity models. The effectiveness of these models was evaluated using the prediction-area (P-A) plot and the normalized density index (Nd). The Nd values for all models exceed one, indicating their effectiveness in integrating exploration evidence to delineate potential targets. However, the DAE-based experiment improved the prediction rate of RF and reduced systematic uncertainties. The proposed methodology was shown to be a robust approach to enhance the relevance of mineral prospectivity mapping, and it may possess the potential to predict new porphyry copper exploration targets in analogous mineral systems.