Deep Forest Modeling: An Interpretable Deep Learning Method for Mineral Prospectivity Mapping

Abstract

Accurate mineral prediction is crucial for reducing costs and uncertainties in mineral discovery and extraction. The use of artificial intelligence and big data has advanced mineral prediction into intelligent forecasting. Machine learning methods have shown significant promise in enhancing outcomes. Currently, neural network‐based approaches dominate deep learning (DL), but they lack interpretability and have high modeling complexity, making them less effective for complex problems and time‐consuming. Deep Forest, an innovative DL paradigm, addresses these issues by dynamically adjusting complexity and providing importance assessments for predictive factors. This study focuses on the North American Cordillera, known for its rich geological data and potential for porphyry copper deposits (PCDs). Predictions are made using Deep Forest with factors like Euclidean distance between faults and magmatic rock, fault line density, gravity anomalies, and stream‐sediment geochemical data. Deep neural networks, random forest, convolutional neural networks, transformer model and graph convolutional networks are also used for comparison. Deep Forest shows high performance and can avoid the black box problem of DL without relying on other tools in DL, providing a new perspective for the development and application of other non‐neural network DL models for mineral prediction. Feature importance analysis shows that geological structure and magmatism significantly influence PCD prediction. Elevated levels of elements like Al, Co, and Cr in stream sediments help identify mineralization‐related alterations. These findings underscore Deep Forest's capability to accurately and efficiently guide mineral exploration, highlighting its potential as a promising approach for mineral prospecting.