Accurate 3D models of deposit lithology are fundamental to enabling orebody understanding; however, these models may be challenging to create directly from the geologists’ lithological logs. DRIVER’s machine learning-supported methods may be used to automatically identify and model lithological units in 3D, providing an alternative method to validate and/or improve the existing geological model.
The example problem dataset hosts a suite of narrow, mafic dikes that cross-cut Sn-ore-bearing sedimentary rocks from the Achmmach tin (Sn) breccia deposit in northeastern Morocco. These dikes were initially modelled implicitly using hand-drawn wireframes that encase the occurrence of mafic intervals obtained from the geologists’ logs.
Figure 1: View looking north at the eastern portion of the Achmmach Sn deposit showing steeply dipping sub-parallel mafic dikes modelled using an implicit (traditional) technique
The 3D distribution of the dikes can be identified through elevated levels of elements that are commonly enhanced in mafic lithologies: P (ppm), Ca (%), Mg (%), Mn (ppm), Na (%) and Ti (%). However, the practical use of these numeric geochemical data for delineating lithological domains would typically require assessment of the geostatistical properties of each element, interpolation of their values into block models, and then manual assessment of valid cut-off concentrations that effectively divide the intrusive and host rock lithologies.
Figure 2 – Composited drilling assays (5 m interval mid-points) showing the concentrations of P (ppm), Ti (ppm), Mg (%) and As (ppm) in the Achmmach deposit
DRIVER provided two independent methods of automatic, machine learning-based modelling that can be used to quickly construct 3D model hypotheses and refine the distribution of the lithological domains.
Method 1: element-by-element modelling and overlap analysis
DRIVER can be used to quickly assess the anisotropy (i.e., directional continuity) of each of the measured elements that define the distribution of intrusive rocks. To begin this process, open the “New global anisotropy estimation” tool and multi-select the assays of interest (e.g., P, Ca, Mg, Mn, Na and Ti). For reference, we will also select As, as it is an element that is not particularly enriched in the dikes, but rather follows a gently east-dipping trend. Before clicking “Submit,” we can verify that the default maximum and minimum feature-length search parameters (347 m and 4 m, respectively) are appropriate (i.e., they broadly capture the area of interest and the maximum length and widths (scale) of features we are trying to model). The results indicate narrow, elongated and moderately inclined ellipsoids for each of the assays enriched within the mafic dike zones. For As, however, the ellipsoid is broad, trends east-west, and dips gently east.
Figure 3 – Anisotropy ellipsoid (left), stereonet (upper right) and variogram (lower right) for P (ppm) showing the orientation and strength of directional continuity
These anisotropy ellipsoids can now be used to inform 3D block model interpolations that respect the true orientation and magnitude of directional continuity of each measured attribute. To create the block models, we open the “New ordinary kriging” tool, multi-select each of the anisotropy objects we previously created, and click “Submit.”
Figure 4: East-west slice through block models for P (ppm), Ti (%) and As (ppm) (left to right) showing the interpolated concentrations of each element
DRIVER also leverages machine learning for automatic clustering of the block model objects and the identification of geochemical anomaly zones. This algorithm is highly flexible and assesses the correlations between individual blocks to define clusters that are statistically anomalous. Block clustering can be applied to multiple attributes simultaneously using the “Auto-generate zones” tool and multi-selecting each block model object. The “Auto-generate zones” function offers various statistical methods to manipulate the results (e.g., minimum concentration threshold, normal score transformation). For this exploratory analysis, the default parameters are adequate.
Figure 5 – Automatically identified 3D volumes for anomalous P (ppm), Ti (%), Mg (%) and As (ppm) (left to right)
The results indicate 3D anomaly zones of P, Ca, Mg, Mn, Na and Ti that are enriched in four NW-dipping, 10-30 m-wide bands. These individual elements are all closely related, both spatially and chemically. Conversely, the identified As anomaly zone forms a broad, east-dipping mass that is apparently unrelated to the lithological boundary between the mafic intrusive and the pelitic host rocks.
A DRIVER-generated mesh intersection (overlap) between P, Mg and Ti forms a close correlation with the traditional mafic dike model.
Figure 6 – Original distribution of modelled mafic dikes (left) and DRIVER multi-attribute overlap zones (P+Mg+Ti – middle) and (P+Mg+Ti+Ca+Mn – right) from machine learning generated anomaly wireframes
Method 2: unsupervised clustering (K-Means) and locally varying anisotropy (LVA) analysis
Method one (above) is a simple procedure that uses the individual geochemical elements as 3D proxies for lithology; however, there are several important drawbacks to this method that we can improve on. Firstly, it relies on singular geochemical elements to be relatively enriched in the dikes, and does not natively account for how these elements may be distributed within the other rock units making up the deposit. As users, we needed to choose and test several elements of importance to define the dikes (e.g., P, Mg, Ti). Secondly, the approach assumed geometric stationarity; in other words, the directions controlling the continuity of each element (anisotropy) were fixed throughout the entire deposit area.
The first problem could be addressed by utilizing an unsupervised multi-dimensional clustering method on the input drillhole data. The K-means method is a popular choice in geoscience data analysis, and we can employ it prior to the data being uploaded into the DRIVER platform (details of this analysis will be covered in a future blog post). The results of K-means on the multi-variate drilling information provide ‘clusters of data’ that are often analogous to lithological or alteration assemblages.
Figure 7: K-Means cluster analysis (colormap showing calculated distances from cluster DP1 center) for the Achmmach deposit
The K-means analysis is effective at grouping the mafic lithologies (Figure 7); however, the results indicate that there is complexity in the dike orientations throughout the deposit. In the central deposit area, the dikes appear to be oriented at ~240/30, while the dikes in the eastern area are oriented at 210/50.
This structural complexity may be accommodated using DRIVER’s locally varying anisotropy (LVA) analysis tool. By restricting the DRIVER anisotropy analysis procedure to smaller, moving windows, the algorithm can effectively map the primary controlling directions as they change locally across the deposit. Like method one, these LVA fields may be used to guide more accurate interpolations (block models), the results of which can be clustered to identify 3D clusters of blocks that are spatially and chemically anomalous (Figure 8).
Figure 8: View looking down on Achmmach deposit showing a horizontal slice through A) the DRIVER locally varying anisotropy field for K-means cluster results acting as a proxy for mafic intrusive, B) interpolation showing 3D block model of predicted association with K-means cluster DP1 across the deposit, and C) DRIVER machine-learning based zone anomalies, a final 3D volume proxy for the mafic dikes
In the next blog post we will cover deposit domaining, and how 3D shapes may be uploaded into DRIVER and used to refine the model results.