Boundary classification for automated geological modelling.

The use of machine learning via Gaussian processes (GPs) has been investigated to automatically detect marker shale bands in natural gamma logs from the banded iron formation (BIF) of the Marra Mamba sequence of the Hamersley Province (Western Australia). Once shale band boundaries have been located, chemical assays from exploration drill holes are used to find the exact boundaries of interest. These divide the drill hole stratigraphy into regions of different mineralogy, each of which display their own distinct correlations between the main elements and oxides (Fe, SiO2 and Al2O3). Iron ore shows a negative correlation between Fe and Al2O3, but in BIF there is a positive correlation between these species. Similarly, SiO2 and Al2O3 have a positive correlation in the shales and ore but a negative correlation in the BIF. Correlations obtained within ore-, BIF- and shale dominated regions are therefore better than those obtained using the entire log and can be used to improve the results obtained when modelling.
The use of machine learning via Gaussian processes (GPs) has been investigated to automatically detect marker shale bands in natural gamma logs from the banded iron formation (BIF) of the Marra Mamba sequence of the Hamersley Province (Western Australia). Once shale band boundaries have been located, chemical assays from exploration drill holes are used to find the exact boundaries of interest. These divide the drill hole stratigraphy into regions of different mineralogy, each of which display their own distinct correlations between the main elements and oxides (Fe, SiO2 and Al2O3). Iron ore shows a negative correlation between Fe and Al2O3, but in BIF there is a positive correlation between these species. Similarly, SiO2 and Al2O3 have a positive correlation in the shales and ore but a negative correlation in the BIF. Correlations obtained within ore-, BIF- and shale dominated regions are therefore better than those obtained using the entire log and can be used to improve the results obtained when modelling.