Abstract
Mineral prospecting mapping (MPM) is necessary for uncovering potential areas for resource exploration and development in the reconnaissance and prospecting stages. However, traditional mapping approaches often suffer from inherent uncertainties due to factors like data quality, geological complexities, and subjective interpretations. This research introduces a novel deep learning framework for MPM using an Artificial Neural Network (ANN) architecture with L1 regularization inspired by SVMs. The approach aims to reduce uncertainty in MPM By harnessing cutting-edge developments in deep learning. It utilizes an MLP architecture with L1 regularization to learn complex patterns from geoscience data and prevent overfitting. The study applies Regularized Deep Learning to create predictive models for copper mineralization prospectivity in the Sardouyeh District, Kerman, Iran. For preparation of initial outputs, we utilized multi-element geochemical patterns obtained through Principal Component Analysis (PCA), mineralization-related geological-structural layers and hydrothermal alteration evidence from the study area, which were transformed into mappable targeting criteria. Additionally, 39 known Cu-porphyry deposits/occurrences and 39 randomly chosen non-prospect locations were used as target variables for model training. The MPM models were evaluated using ROC, F1-score, confusion matrix, and precision metrics. The Regularized MLP model showed superior prediction accuracy over traditional ANN algorithms, achieving 96% accuracy, 95% recall, 97% precision, 96% F1 score, and 99% AUC for Cu-porphyry deposits. This study highlights the importance of advanced machine learning techniques, specifically Regularized Deep Learning, for optimizing hyperparameters, reducing uncertainty, increasing precision, achieving high accuracy, enhancing mineral exploration efficiency, and addressing overfitting challenges in traditional ANNs.