What happens when you put cutting-edge machine learning up against traditional techniques for geological modelling?
You get a geological model, in less than 50 minutes.
Machine learning vs traditional geological modelling
The Lisheen Mine, a carbonate hosted zinc-lead project located in the Irish midlands in County Tipperary was in full operation from 1999 until 2015, when it ceased production. During that period the site used Maptek™ Vulcan™ software for underground survey, geological modelling, variography, resource estimation and mine planning.
We used the Lisheen data to test how our machine learning tool performed compared to the traditional geological modelling techniques that were used during mining.
What sort of time-savings would we see? Would we see a close match between the geological models?
What is machine learning and why do we need it?
Domaining, resource estimation and grade control are critical mining processes which involve complex, specialist skills and a significant investment in time and resources. DomainMCF, developed within the new Maptek framework for cloud computing, uses machine learning techniques to resolve these challenges faster and more efficiently than has been traditionally possible.
Let’s put the complexity of resource estimation in context: setting up a single variable in a single geological domain using the inverse distance technique could entail up to 190 decisions.
Applying ordinary kriging to a single variable/domain can take up to 170 decisions in preparation alone to determine appropriate parameter settings, followed by a further potential 220 decisions to carry out the kriging estimate.
A deposit with a single commodity of economic interest within a single uniform geology is rare. For example, one Maptek customer has approximately 900 variable/geological domain combinations. Applying kriging to the entire deposit requires about 35,000 decisions!
Machine learning seeks to unlock new algorithms to resolve these challenges faster and more efficiently. The technology allows machines to solve complex problems even when using a data set that is diverse, unstructured and inter-connected. The more deep learning algorithms learn, the better they perform and as a result mining companies gain a greater understanding of their deposit.
The Lisheen mine and the scope of our challenge
2700 surface and 4670 underground drillholes (see Figure 1) were drilled at Lisheen, each one logged in detail by mine geologists. Rock types were classified into 50 different codes during logging and these were used for subsequent domain modelling.
We don’t know precisely how long it took to generate the wireframes to represent these geological codes, but it would be safe to assume that multiple weeks were required to model the complex interaction of some of these units.
How fast could machine learning process and model our data?
Lisheen data was entered into the machine learning engine for processing along with dimensions for the resulting block model, into which domains are directly written.
No other geological interpretations, aside from the rock codes in the drill base were used for processing. A parent block size of 12x12x4m was used with subcell sizes down to 3x3x1m for the subhorizontal orebodies.
Within 50 minutes the block model was complete, flagged with domains generated for all 50 rock codes.
So far so good, but how would the model made by our machine learning system compare with the original Vulcan model?
The visual and wireframes showed a very close match and there was only a 4% difference in the ore volumes between both methods. We’re rather happy with that.
In conclusion it took DomainMCF less than 50 minutes to flag geological domains. The same process would have taken up to multiple weeks using traditional methods.
Marketing Manager EMEA
March 7, 2020
For additional information about Maptek, including use of the Maptek logo, product images and reproduction of case studies, please direct inquiries to Global Marketing Communications Manager firstname.lastname@example.org