Abstract

Part of basic well log analysis involves the segmentation of petrophysical logs with formation tops within a log. This segmentation is frequently extended to the interpretation of nearby logs in a petroleum basin. Generally speaking, this type of correlation of well logs is subjective. This can be a challenge in areas where data is complicated and opinions begin to diverge. It is reasonable to speculate that opinions diverge because, in parts of some logs, it can be difficult for the human brain to discern structure in complicated petrophysical signals. This subjectivity leads to the injection of additional noise into the data set. For example, different geologists may place a formation top in several different locations on the same petrophysical log.

The objective of this research is to provide protocols for segmenting petrophysical well logs using methods borrowed and adapted from nonlinear dynamics and also to show a practical application in n-dimensional data clustering and supervised learning. This study uses a change in fractal dimension of tool response to quantitatively locate lithological transitions and then use those transitions to train a machine to identify different subsurface objects. The ultimate goal is to provide another tool for geoscientists to use on hard to interpret petrophysical logs using real world data backed up by synthetic models to inform intuition. To demonstrate real world applications of the method, this study utilizes petrophysical logs from various basins and then applies this by using pattern recognition to identify an unconventional formation of interest in the Cooper Basin, Australia.

The method employed in this study for the characterization of the fractal geometry of petrophysical logs shows promise. The application of the method is tricky at times because its effectiveness is very dependent on the choice of a windowing parameter. However, when the parameters are appropriately selected the method is frequently able to detect a change in the structure of the petrophysical signal. Moreover, using a change in fractal dimension to characterize tool response enables the user to detect changes in tool response that might have otherwise gone unnoticed. Using a change in fractal dimension appears to isolate objects that traditional clustering techniques are so far unable to differentiate.