Duy Thong Kieu

Testing cluster analysis on combined petrophysical and geochemical data for rock mass classification

New drilling, measurement-while-drilling and top-of-hole sensing technologies are being developed to overcome the challenges of exploration for new mineral deposits under deep cover. These methods will provide continuous, near-real time data collection from every drillhole in the future. Consequently, there will be a need for efficient methods of analysing and interpreting this data stream to complement the exploration strategy. We demonstrate the usefulness of cluster analysis for rapid, automated rock mass classification, and the impact of selecting different subsets of the available data on the classification results. Our study shows that only a few measurements are needed to broadly domain the intersected rock mass and highlights the importance of selecting correct input data depending on the purpose of the classification. Our analysis also indicates the potential of identifying textural and rock mechanical properties from petrophysical measurements via cluster analysis. We demonstrate the usefulness of fuzzy clustering for semi-automated rock mass classification and the importance of data preconditioning prior to analysis. We also show the impact of choosing different subsets and combinations of the available data on the classification result.

Estimation of P-wave velocity from other borehole data

Summary P-wave velocities are a key parameter for seismic processing and the absence of this parameter reduces the robustness of the images from very expensive seismic surveys. The P-wave velocities in an area are particular to the area, as the P-wave velocity depends on many factors and varies with geological conditions. Hence, using a localized model predicts P-wave velocity better than the application of a generic model for the entire dataset. In this work, we utilized fuzzy c-means (FCM) clustering to build a “fuzzy” relationship that estimates Vp. Our method was tested on a dataset from the Kevitsa Ni-Cu-PGE deposit in northern Finland. The borehole data comprises P-wave velocity, density, natural gamma, magnetic susceptibility, resistivity and assay data of Ni of six boreholes. In this area, there are many boreholes, but very few have P-wave velocity logged or the data is corrupted by tool limitations. Therefore, it is beneficial to predict the velocity from other data to help seismic processing. In order to demonstrate the robustness of our program, we used the data from five holes for training and one hole for Vp testing. The results show that our method can reasonably estimate P-wave velocity from other borehole data.

Prediction of sonic velocities from other borehole data: An example from the Kevitsa mine site, northern Finland: Prediction of sonic velocities

P-wave and S-wave velocities are vital parameters for the processing of seismic data and may be useful for geotechnical studies used in mine planning if such data were collected more often. Seismic velocity data from boreholes increase the robustness and accuracy of the images obtained by relatively costly seismic surface reflection surveys. However, sonic logs are rarely acquired in boreholes in-and-near base metal and precious metal mineral deposits until a seismic survey is planned, and only a few new holes are typically logged because the many hundreds of holes previously drilled are no longer accessible. If there are any pre-existing petrophysical log data, then the data are likely to consist of density, magnetic susceptibility, resistivity and natural gamma logs. Thus, it would be of great benefit to be able to predict the velocities from other data that is more readily available. In this work, we utilize fuzzy c-means clustering to build a “fuzzy” relationship between sonic velocities and other petrophysical borehole data to predict P-wave and S-wave velocity. If boreholes with sonic data intersect most of the important geological units in the area of interest, then the cluster model developed may be applied to other boreholes that do not have sonic data, but do have other petrophysical data to be used for predicting the sonic logs. These predicted sonic logs may then be used to create a three-dimensional volume of velocity with greater detail than would otherwise be created by the interpolation of measured sonic data from sparsely located holes. Our methodology was tested on a dataset from the Kevitsa Ni-Cu-PGE deposit in northern Finland. The dataset includes five boreholes with wireline logs of Pwave velocity, S-wave velocity, density, natural gamma, magnetic susceptibility and resistivity that were used for cluster analysis. The best combination of input data for the training section was chosen by trial and error, but differences in themisfit between the various training datasets were not particularly significant. Our results show that the fuzzy c-means method can predict sonic velocities from other borehole data very well, and the fuzzy c-meansmethod works better than using multiple linear-regression fitting. The predicted P-wave velocity data are of sufficient quality to robustly add low-frequency information for seismic impedance inversion and should provide better velocity models for accurate depth conversion of seismic reflection data.

Fuzzy clustering constrained magnetelluric inversion - Case study over the Kevitsa Ultramafic Intusion, Northern Finland

Inverse magnetotelluric (MT) problems are non-unique and smoothing criteria are typically added to choose the “best” model. However, the process often produces an unrealistic geological model. In reality the subsurface geology is differentiated by distinct rock units that are often better defined by boundaries rather diffuse or smooth boundaries. We present the application of fuzzy clustering as an added constraint within the inversion process to guide model updates toward earth models that are “blocky”, and thus resemble geological units. Fuzzy clustering divides the simulated model into clusters based on the similarity of model features. Moreover, fuzzy clustering naturally enables the inclusion of additional prior information in the inversion process, such as petrophysical information from borehole data. The inclusion of this information produces geo-electrical distributions that are more representative of the true rock units. This is demonstrated through the case study of the Kevitsa Ni-Cu-PGE deposit, northern Finland. The inversion can detect the ore zones and carbonaceous phyllite relating to the conductive zones. The inverted cluster generated model is compare better with borehole data than other approaches.

Seismic Impedance Inversion with Petrophysical Constraints via the Fuzzy Cluster Method

Seismic impedance inversion produces results that should be better for geological interpretation. However, seismic impedance inversion in mineral exploration normally suffers from poor signal-to-noise, and a lack of well control normally assumed for the process. To counter these problems we have developed an approach that exploits the fact that the geology in these environments often has fewer distinct geological units so we can restrict the number of physical parameters possible. A model-based seismic impedance inversion method using fuzzy c-means clustering to constrain inversion with petrophysical information has been developed. Using synthetic examples, we show that our method effectively recovers the true model even when the data is strongly contaminated by noise. This method is applied to seismic data from a US gold mining district and the results are reasonably consistent with well log data. The impedance images provide a better basis for geological interpretation than reflection images alone.

Classification of Geochemical and Petrophysical Data by Using Classification of Geochemical and Petrophysical Data by Using Fuzzy Clustering

In this study, the fuzzy c-mean clustering method was used in an unsupervised manner to automatically classify the different lithologies present at the Hillside prospect (Yorke Penninsula, SA). The algorithm was applied to various combinations of petrophysical and geochemical data to identify the combination that returned the most accurate result and the smallest combination that provides a nearly identical success as the best. We show that by using a combination of geochemical and petrophysical data the likelihood of a correct classification increases by 5% compared to analysing only geochemical data, and by over 20% compared to analysing only petrophysical data. However, using a few common elements and a few petrophysical values we can achieve almost the same success rate as the best result. Improvements in pre-treatment and conditioning of the data should allow the fuzzy cluster algorithm yield even better results. In addition to showing that combining petrophysical and elemental analysis is more robust, we demonstrate that if we could add some targeted elemental analysis to logging while drilling (LWD) then robust automated lithological logging becomes feasible.