Introducing HORIZON: Pioneering Critical Metals Exploration with Deep Learning and Databricks
Introduction
Durendal Resources has developed HORIZON, an advanced Deep Neural Network designed for mineral exploration. Named the High-Resolution Ore Investigation Network, HORIZON can accurately classify mineral occurrences without relying on surface geochemical or direct detection methods, making it a valuable tool for exploring areas with cover. Utilizing deep learning, HORIZON analyzes extensive geological and exploration data to reveal hidden patterns Deng et al. (2021).
Mineral exploration has always been a difficult and complex endeavor, particularly in regions where mineral deposits are not directly observable on surface and are located beneath a large amount of cover. González-Álvarez et al. (2020) Traditional methods often prove to be costly and ineffective in these scenarios. HORIZON has been initially trained on the Lachlan Orogen in New South Wales.
The Lachlan Orogen
Stretching from northeastern Tasmania to central and eastern New South Wales, contains various significant mineral deposit types such as orogenic gold (e.g. Stawell, Ballarat, Bendigo), volcanic-hosted massive sulfide (e.g. Woodlawn, Currawong), sediment-hosted Cu-Au (e.g. Cobar Basin deposits), porphyry Au-Cu (e.g. Cadia, Parkes, Cowal) and granite-related Sn (e.g. Ardlethan, Beechworth). Hough et al. (2006).
Surface exploration in New South Wales is unlikely to yield significant undiscovered metalliferous mineral deposits. Therefore, a novel approach for exploration targeting is essential to discover deposits at depth and beneath the surface cover.
The Opportunity in New South Wales
New South Wales is renowned for its gold and copper resources. Despite a long mining history, vast areas remain unexplored. Notably, most drill holes in this region are shallow, with only 7.4% exceeding 150 meters in depth, indicating significant untapped potential.
Data from MinEx Consulting shows that deeper gold discoveries are often much larger and more valuable than near-surface deposits. Additionally, many exploration efforts focus on areas with little cover, leaving vast regions open for new discoveries using advanced technologies like HORIZON.
Challenges of Traditional Exploration Methods
Traditional exploration relies on detecting and sampling exposed rock formations, which works well in areas with visible outcrops. However, this method is less effective when mineralization is hidden beneath soil or weathered rock, leading to expensive and time-consuming drilling and sampling.
The Potential of Machine Learning in Mineral Exploration
Machine learning (ML) and artificial intelligence (AI) offer new ways to overcome these challenges. By analyzing large datasets, Deep Neural Networks (DNNs) can identify subtle patterns indicating mineral systems, improving prediction accuracy and reducing the need for extensive fieldwork.
Integration with Databricks
Databricks is a key part of our technology stack. This data analytics platform enables efficient data processing, machine learning, and collaborative analytics. Databricks integrates various data types, allowing advanced preprocessing techniques that optimize neural network performance. Its collaborative features also enable seamless teamwork among geologists, data scientists, and ML engineers, speeding up model development and fine-tuning.
When choosing between Databricks and Azure ML for advanced data analytics and machine learning, Databricks stands out due to its unified platform that seamlessly integrates data engineering, data science, and machine learning workflows. Built on Apache Spark, Databricks offers unparalleled scalability, performance, and collaboration capabilities. Its interactive notebooks support multiple languages, enhancing flexibility for diverse teams. The platform’s robust optimization features and automated cluster management reduce operational overhead, allowing teams to focus on innovation. Additionally, Databricks’ deep integration with other Azure services ensures a smooth and efficient deployment process, making it the preferred choice for comprehensive, large-scale data projects.
The Lachlan Orogen: Case Study
The Central Lachlan Orogen, with its significant cover and mineral potential, is ideal for advanced ML techniques. Traditional methods have mainly focused on areas with exposed rock, leaving vast covered regions unexplored.
Data Collection, Preparation, and Feature Engineering
HORIZON uses detailed geological, geophysical, and structural data, including fault characteristics, magnetic and gravity data, and stratigraphic features. The data is spatially indexed using a proprietary algorithm to maintain geographical consistency, ensuring a comprehensive understanding of the region.
Model Training and Evaluation
HORIZON was trained on this detailed data, divided into training, validation, and test sets to ensure effective learning and generalization. Hyperparameter tuning with Optuna improved the model’s architecture and predictive accuracy.
Model Performance
HORIZON achieves high performance, with an F1-score of 0.95 for identifying mineral deposits and 1.00 for correctly identifying areas without deposits, making it a reliable tool for exploration.
Deployment and Insights
Once trained, HORIZON generates mineral classification maps for the Central Lachlan Orogen, providing valuable insights into promising exploration targets and guiding field operations. Integration with Databricks allows efficient data processing and rapid generation of actionable insights.
Conclusion
HORIZON represents the first step in a revolutionary suite of models designed to transform mineral exploration worldwide. While significant progress has been made, further development is needed to create models that can make highly accurate predictions with minimal geoscience data. This ongoing work promises to enhance the efficiency and effectiveness of mineral exploration, uncovering hidden resources and driving the industry forward.
Acknowledgements
Aakash Kolte, Data Scientist; Kartik Pardeshi, Machine Learning Engineer; Sonu Kumar, Machine Learning Engineer; Vasu Katravath, Data Scientist; William Vallier, CTO; Paul Dale, VP Global Exploration; Mark Armstrong, Rob Reid, Stefan Zanon, Maarten Sundman, Jatin Kumar, Manoj Yadav, Mounika Adabala, Shubham Narwade
References
Deng, H., Yang, Z., Jin, C., Yu, S., Xiao, K., & Mao, X. (2021, September 2). Learning 3D Mineral Prospectivity from 3D Geological Models with Convolutional Neural Networks: Application to a Structure-controlled Hydrothermal Gold Deposit. https://arxiv.org/pdf/2109.00756v1.pdf
González-Álvarez, I., Gonçalves, M A., & Carranza, E J M. (2020, November 1). Introduction to the Special Issue Challenges for mineral exploration in the 21st century: Targeting mineral deposits under cover. Elsevier BV, 126, 103785–103785. https://doi.org/10.1016/j.oregeorev.2020.103785
Hough, M., Bierlein, F P., & Wilde, A. (2006, July 4). A review of the metallogeny and tectonics of the Lachlan Orogen. Springer Science+Business Media, 42(5), 435–448. https://doi.org/10.1007/s00126-006-0073-7
