Machine learning for grade control

A study of a haematite and magnetite deposit submitted for the Maptek Geology Challenge found that machine learning holds promise for grade control.

SIMEC Mining produces iron ore from active operations in the Middleback Ranges, about 50 km west of Whyalla in South Australia. Mineralisation occurs predominantly as haematite and magnetite hosted in large-scale banded iron formations. These rocks are complexly folded by at least three major deformation events and are intruded by multiple phases of mafic–intermediate dykes and granite intrusions.  

Prompted to enter the Maptek Geology Challenge, SIMEC Geology Superintendent Ed Lynch was interested in the opportunity to test Maptek DomainMCF on these highly complex structures.

The annual Geology Challenge inspires geologists to engage with new approaches and redefine the boundaries of geological modelling. 

DomainMCF uses machine learning to rapidly produce domained block models—an objective and repeatable approach potentially saving significant time and cost. 

Lynch set up various scenarios using both real and simulated grade control drilling data to test DomainMCF and assess possibilities for improving current modelling and grade control practices. One of these scenarios is discussed here.

Magnetite domains

For magnetite grade control, 6m composite samples taken at an approximate 10mx20m spacing are analysed by Davis Tube Recovery (DTR) to ascertain recoverable metal and associated concentrate grades. Material is segregated based on the DTR results—primarily mass recovery, proportion of silica in the concentrate product, and proportion of sulphur in the original sample. 

There is strong geological control on these material classifications as they correlate to stratigraphic units, with cut-off criteria determined through extensive lithogeochemical classification. These material categories form the basis for the domains used as inputs to the DomainMCF modelling process.

To validate the accuracy of the DomainMCF models, block models were developed independently through existing grade control practices, using manual interpretation and grade estimation by ordinary kriging methods. 

Those block models produced during ongoing operations are considered the closest possible representation of actual domain volumes realised during mining and so were used as benchmarks to reconcile against and assess the accuracy of DomainMCF models.

Multi-bench grade control 

Lynch considered three recently mined benches with 1601 magnetite grade control samples. Two DomainMCF models were produced using the same machine trained on sample data domained based on magnetite grade control cut-offs. 

One model had block dimensions half the sample spacing and length (5mx10mx3m) and another had the same block dimensions used in the grade control model (25mx25mx4m), with sub-blocking enabled to one-quarter of the parent block dimensions at domain boundaries. The second model was run using the previously trained machine to test for repeatability. 

Each model was completed within a total compute time of approximately two minutes, and the models were validated visually against the input samples. The resultant models run at different block sizes using the same machine produced identical domain volumes.

The DomainMCF models produced at smaller block sizes demonstrated a more geological appearance by removing polygonal or jagged edges without impacting the general trends and volumes of modelled domains.

Lynch concluded that there is no apparent drawback to using smaller blocks apart from longer computation times. Given the efficiency of DomainMCF compared to traditional modelling methods, computation times even at smaller block sizes are not a concern.

Accuracy of the DomainMCF models was assessed largely by reconciling the domain volumes against the volumes of like domains in grade control models. In this case total ore volumes modelled by DomainMCF differed by only 2.1% from actual grade control models. 

This demonstrates that DomainMCF can be successfully applied to drilling data at resolutions typical of wider spaced reverse circulation drilling grade control programs and could replace or supplement current methods. 

Validation 

While the DomainMCF process removes the subjectivity of human geological interpretation, the validation stage requires geologists with suitable expertise to determine whether the model satisfactorily represents all of the available data and underlying geology. 

For example, SIMEC mine geologists isolate material containing sulphur, which is considered a deleterious element. Although only 1.3% of the total volume, it is important in the grade control model. In this case, relatively very few samples of this material were available, leading to misleading results.

Models informed by a larger number of samples at tighter spacings were generally more representative of the input data and reconciled more closely to grade control.

Spatially the multi-bench models closely represent the input data in both plan and section views with the differences in volumes produced by DomainMCF compared to the actual grade control not visually obvious.

Conclusion 

This study demonstrated that DomainMCF can efficiently produce domained models that represent the input data to an acceptable level of accuracy for grade control. Models are repeatable and produced in a fraction of the time through traditional methods—minutes and hours rather than days or months. 

DomainMCF presents a valuable tool that can drive improvements and efficiency in grade control processes which then feed into mine planning and operational decision making.

Thanks to 
Ed Lynch, Geology Superintendent
Matthew Peacock, Chief Geologist
Colin Badenhorst, Principal Resource Geologist
SIMEC Mining