Targeting Core Sampling with Machine Learning: Case Study from the Springbok Sandstone, Surat Basin

Abstract

We show how clustering algorithms can ensure that the core intervals that are pertinent to specific objectives of a sampling campaign are actually sampled. We also show how clusters can be validated prior to sampling with auxiliary data not used for the cluster analysis. We chose to target our core sampling to ensure that both clay poor and clay rich intervals of the Springbok Sandstone are sampled. The clay phases in the Jurassic Springbok Sandstone generally do not exhibit a prominent gamma ray signature and are therefore poorly defined in wireline logs. Similar, hydrogeological properties of the Springbok Sandstone are not well defined through wireline logs. This introduces uncertainty to groundwater models of the Springbok Sandstone. Hence, a better understanding of the clay distribution is thought to be a key to improve the definition of the hydrogeological properties of the Springbok Sandstone. We applied our sample targeting approach to five study wells from the Surat Basin in Queensland. We tailored the application of the cluster analysis to our working hypothesis that the variability of hydrogeological properties of the Springbok Sandstone is controlled by the presence and type of clays, rather than compaction. This informed our choice of wireline logs to include in the clustering (nuclear logs) and of logs to be used for control purpose (resistivity logs, spontaneous potential). We show that identification of five clusters was the most useful number towards our sampling objectives. This allowed for example to exclude coal and siderite layers from sampling for clay analysis and to focus on the differentiation of the clastic sediments in the formation. Further, we show that certain clusters correlate with resistivity and spontaneous potential log signatures. The correlation between the categorical clusters based on nuclear logs and continuous wireline logs not used in the cluster analysis allowed us to interpret the meaning of the clusters in the context of our project and target our sampling to ensure that all clusters are represented in our sample set.