Machine Learning models help geologists visualize the features of rocks that would otherwise remain invisible deep underground.

In our last post, we discussed how Artificial Intelligence can improve mining resource estimation in cost, time, and prediction accuracy. Here we drill deeper into how data scientists and geologist make machines smart enough to see more of what’s underground with less drilling.

The problem with drilling is that, while it provides relatively reliable ore samples for analysis and validation, it’s expensive, slow, and cumbersome. Core drilling to obtain physical samples for lab analysis, data consolidation and reconciliation, validation, modeling, and publishing can take up to 17 weeks.

Machine Learning that correlates massive data can do it in one. And predict ore qualities with very high confidence, cheaper, and with less drilling. So, how do you make a machine smart?

Teachable Machines Instead of Core Drilling

Artificial Intelligence (AI) is a catchall term that includes Machine Learning, neural networks, deep learning, computer vision and natural language processing. AI generally refers to the broad idea that machines can execute tasks “intelligently.”

Machine Learning refers to the concept that machines can learn and adapt through experience. A neural network is a computing system modeled after the human brain. Deep learning uses huge neural networks with many layers of processing networks. Computer vision uses pattern recognition and deep learning to recognize what’s in a picture or video. Natural Language Processing allows computers to analyze, understand and generate human language and speech.

Data scientists often assemble and combine different AI approaches in various ways to quickly interpret oceans of data no single person or team can manage alone. It’s about pattern recognition across disparate databases from wide-ranging sources: government geological records, site equipment sensors, borehole samples, satellite imagery, magnetic field readings, topographical maps, and field reports.

Machine Learning models for resource estimation interpret correlations among geoscience data to create geological and resource models that rely less on expensive, time-consuming core drilling operations. Faster and less expensive non-core drilling can be used to provide the bulk of physical samples for quick geophysical logging, but weeks-long lab analysis of core drill samples isn’t as necessary with a Machine Learning model.

Building a Smart Machine

To build datasets and define the right models, data scientists and geologists have several Machine Learning approaches they can use. Examples include approaches known as random forest (DRF), extreme random forest (XRT), generalized linear model (GLM), gradient boosting machine (GBM) and Deep Learning. Ranges of parameters must be configured, and the best model in each class selected to build a “stacked ensemble.”

The Machine Learning models, configured with geological databases and algorithms designed to adapt to new experiences and data updates (learning), take days to build rather than months using current core drilling methods.

These methods are being tested and the results are remarkable. In 2020 a team of researchers known as DeepSightX combined government magnetic and gravity data with deep neural network AI analysis to develop Machine Learning approaches that helped geologists visualize the features of rocks that would otherwise remain invisible deep underground in remote South Australia. In another recent pilot, resource geologists and data scientists developed Machine Learning models with historical geophysical data to establish orebody properties for non-core drilling holes. The project generated in a matter of days predictions of critical properties and lithology with a 95 percent confidence rating.

In our next installment of this series​​​​​​​, we’ll look at early Machine Learning applications and pilot studies and how they might affect the mining industry in the near term.