���˴�Ƭ

New interpretive maps revealing undiscovered mineral deposits deep underground in remote South Australia have been developed by combining newly released government magnetic and gravity data with deep neural network AI analysis. 

The approach was developed by South Australian team DeepSightX, and received a merit award in the , a global award targeting innovation to identify valuable new mineral deposits in South Australia’s Gawler region. As revealed today, DeepSightX and four other entrants made the finalist list in 2020, with more than 2,200 data scientists and geologists from over 100 countries involved this year. 

DeepSightX is led by , Director in Advanced Reasoning and Learning at the Australian Institute for Machine Learning at the ���˴�Ƭ. 

“Our application demonstrates best practices in deep learning, enabling experts in geophysics to produce maps of geological transitions and structures. Our system estimates valuable underground mineral deposits and underlying mineral and geological structures.” said Professor Shi. 

“What we end up with is a fine-grained map of likely mineral distribution over the Gawler region.” 

“This information can now be applied to make decisions about more detailed investigation of areas of interest,” said Professor Shi. 

The Gawler Challenge is run by the Government of South Australia, and is designed to encourage experts to develop new technologies for fast, highly targeted geological exploration. 

Quest to find high value minerals

Experts believe mineral deposits of high economic value are hidden deep underground in South Australia. But the key question is where to start looking. 

“South Australia has one of the world’s best copper provinces, demonstrated by the early discoveries during 1975 to 2005,” said John Anderson, Principal Consultant at Austrike Resources and DeepSightX team member.  

“But we think there are a large number of copper discoveries still to be made – and perhaps also nickel and gold.” 

Alan Collins, Professor of Earth Sciences at the ���˴�Ƭ and who was not involved in the project, agrees that significant mineral deposits are still to be discovered in South Australia. 

“There are undoubtedly considerable mineral deposits to be found under younger rocks in the Gawler Craton,” he said.

“The discovery of Carrapateena in 2005 is an example of how targeting exploration and targeted government support for exploration can help discover significant resources.” 

Carrapateena is an Oz-Minerals copper-gold project for which average annual production is expected to be 65,000 tonnes of copper and 67,000 ounces of gold over the life of the mine. 

DeepSightX took The Gawler Challenge as an opportunity to evaluate artificial intelligence (AI) applications in combination with local geological expertise to jump-start the next generation of mineral discoveries. 

“The DeepsightX study focused on developing better interpretive maps of the geology not visible at the surface. These initial concepts developed by the team’s geophysicist Matt Zengerer were enhanced by applying AI to new government magnetic and gravity data,” said John Anderson. 

Invisible rocks deep underground

The Gawler region consists of around 440,000 square kilometres north west of Adelaide. The area is well known for its highly valued copper and gold deposits, including iron-oxide-copper-gold deposits at Olympic Dam and Prominent Hill. 

But most of the area is covered with young sediments which obscure the ancient rocks. With all the obvious mineral deposits already identified, geologists are seeking new ways to identify untapped areas of mining potential and bring value to South Australia.

DeepSightX’s machine learning approaches help geologists visualise the features of rocks that would otherwise remain invisible. 

“AI plays a critical role, as it massively shrinks the problem of trying to accurately map the subsurface rocks and mineralisation from years to potentially weeks in specific regions,” said Matt Zengerer, Principal at Gondwana Geoscience and DeepSightX team member. 

“With robust implementation and rapid uptake we may see these tools being used within 6 to 12 months by many companies on a more routine basis.”  

How machine learning works with geology 

Machine learning works by training algorithms with large data sets so that patterns not detectable by humans alone can be identified. In this project, the data used to train the AI describes the magnetic, gravitational and other physical features of Gawler region rocks.

Professor Shi and his team developed six unique machine learning approaches, which they grouped into two major strategies relating to applications in geophysical exploration.  

“Firstly, we demonstrated how DeepSightX can inject deep learning best practices into existing geophysical exploration pipelines,” said Professor Shi.

“Secondly, we applied our expertise in sensor processing in the problem of approximating the mineral and geological subsurface with a scalable data-driven approach. That enables predictions of minerals and geology in vast unknown regions by learning a direct link between cheap and vastly available sensory data to costly and sparsely collected drillhole data.” 

Professor Shi said that in addition to creating a fresh approach to data-driven target hunting and geophysical analysis, their application also demonstrated various methods of robustly testing the models for when they go into production. 

“Everyone promises extracting value with AI, but we present two paradigms that are verifiable by industry experts,” he said. 

“The first approach provides a tool to automate labour-intensive processes, and the second provides a fresh perspective for exploration,” said Professor Shi.

As part of the competition, the Explore SA: Gawler Challenge provided data from an airborne magnetic exploration survey called the Gawler Craton Airborne Survey, conducted in South Australia from 2016 to 2019. The DeepSightX team used this data along with gravitational data and details of known mineral deposits and the physical features of rocks to develop their neural network (a multilayered AI system). 

The DeepSightX team: 

Javen Qinfeng Shi: Australian Institute for Machine Learning (AIML), ���˴�Ƭ 

Adrian Orenstein: AIML, ���˴�Ƭ 

Mahdi Kazemi Moghaddam: AIML, ���˴�Ƭ

Matthew Zengerer: Gondwana Geoscience

Ehsan Abbasnejad: AIML, ���˴�Ƭ

John Anderson: Austrike Resources Pty Ltd 

Hao Zhang: DeeperX

Lingqiao Liu: AIML, ���˴�Ƭ

Anton van den Hengel: AIML, ���˴�Ƭ

Chris Matthews: Institute for Mineral and Energy Resources (IMER), ���˴�Ƭ

Story written by Dr Sarah Keenihan, AIML