TL;DR
This paper introduces a novel application of artificial neural networks (ANNs) for continental-scale mineral exploration, specifically focusing on gold deposits in the Andes. By leveraging a comprehensive GIS database and employing a pruning algorithm, the study demonstrates how ANNs can effectively identify key geological attributes controlling deposit formation, leading to more efficient and data-driven exploration strategies.
What Problem Does This Paper Address?
Traditional mineral exploration relies heavily on geologists' expertise and conceptual models, which can be limited in scope and subjective. While Geographic Information Systems (GIS) have revolutionized data integration, extracting meaningful insights from vast, multi-layered datasets remains a challenge. This paper addresses the need for a more objective, data-driven approach to mineral potential mapping, particularly at a continental scale, where general domain knowledge alone is insufficient to model the complex factors influencing ore deposit formation. The authors aim to overcome the limitations of traditional methods by using ANNs to learn complex spatial relationships and identify crucial predictive attributes from a large GIS database.
Methods
The core of this research lies in the application of Artificial Neural Networks (ANNs), specifically multilayer perceptrons (MLPs), for classification. The ANNs were trained to distinguish between mineral deposit sites and "barren" sites. A key methodological innovation is the use of the Optimal Brain Damage (OBD) algorithm for automatic attribute selection. This pruning technique iteratively removes less influential connections (weights) or neurons within the network. The goals of OBD are manifold: to improve generalization performance by reducing overtraining, to simplify the model for faster operation, and crucially, to extract knowledge by identifying the most relevant attributes and potentially uncovering new rules or combinations of factors influencing mineralization. The network architecture involved 25 input variables, a hidden layer with varying units, and two output units using a softmax activation function to predict the probability of a site being a "deposit" or "barren."
Datasets
The primary dataset used in this study is the GIS Andes, a comprehensive, spatially referenced geographic information system covering the entire Andes Cordillera. This system integrates numerous thematic layers, including:
- Geographic base data (DCW®)
- Digital Elevation Models (DEM)
- SPOT 4 VEGETATION® imagery
- Geological synthesis maps
- Seismic data
- Volcanic data
- Gravimetric data
- Heat flow data
- Geochemistry data
- Ore deposit databases
- Mineralogy, fluid inclusions, and isotope data
For training and evaluation, two specific classes were derived from GIS Andes:
- "Deposit" variable: Comprising 398 recent (2-15 Ma) Andean Au, Ag, Cu, Mo, Sb, Hg, Pb, and Zn-bearing deposits (epithermal-porphyry type), documented as points.
- "Barren" variable: Consisting of 243 arbitrarily selected points from areas well outside known mining districts and where no deposits have been discovered. This class is crucial for the network's learning process to discriminate against actual deposits.
The paper does not introduce new benchmark datasets but rather utilizes an existing, extensive continental-scale GIS.
Key Insights and Ideas
- ANNs as Knowledge Extractors: Beyond mere classification, ANNs, when combined with pruning techniques like OBD, can serve as powerful tools for extracting geological knowledge that might not be explicitly captured in existing conceptual models.
- The Importance of "Barren" Sites: The study highlights that defining and selecting appropriate "barren" sites is critical for effective ANN training. These sites, while lacking geological significance in themselves, are essential for the network to learn the contrast and discriminate between mineralized and non-mineralized areas.
- Attribute Relevance is Context-Dependent: The OBD algorithm revealed that attributes like "shortest distance to fault orientations" and "earthquake density" are more influential for gold mineralization in the Andes than initially assumed, underscoring the need for data-driven attribute selection.
- Surprising Attribute Omissions: The "nature of host rock" (e.g., volcanic lithologies), often considered a primary factor, ranked surprisingly low in importance after pruning, suggesting that other factors might be more dominant in this specific geological context.
Key Results
- Discrimination Performance: The ANNs achieved good discrimination performance for specific metals: 92% for gold, 93% for silver, and 88% for copper. Performance was less effective for metals with fewer training examples (e.g., antimony, mercury).
- Variable Selection: The Optimal Brain Damage algorithm successfully reduced the number of attributes and provided an ordered list of the most influential factors for gold mineralization. The top attributes included:
- Shortest distance to fault orientations (N 135°E to N 157.5°E)
- Earthquake density
- Longitude
- Presence of Cenozoic host rock
- Shortest distance to an active volcano
- Generalization: A multilayer perceptron using all 25 attributes and 10 hidden units achieved an 84% correct classification rate in generalization (Table 1), with 66% of actual deposits correctly identified and 94% of actual barren sites correctly identified.
Limitations and Future Work
- Data Dependency: The performance of data mining and ANNs is heavily dependent on the quality and representativeness of the database. The selection of "barren" sites, while crucial, is acknowledged as an area for ongoing research to ensure they closely resemble deposit sites in location and data description.
- Metal Specificity: The study notes that discrimination performance varies by metal, suggesting that models might need to be tailored for different types of mineralization.
- Future Work: The authors suggest that the approach can be further refined by carefully selecting barren sites that are similar to deposit sites but lack the specific metal of interest. The research also implies that the extracted knowledge from the pruned network could be used to develop more sophisticated predictive mineral maps and potentially uncover new rules for deposit formation. The ongoing research aims to incorporate these refined barren site selection strategies into the network's operation.
Summary updated