ML for Drilling Problem Identification

Geothermal district heating is a major opportunity for decarbonisation, with nearly 40% of all UK energy used for heating and natural gas the predominant heat source. However, achieving sustainable and sufficient heat supply requires drilling into permeable strata at depths below the Earth's surface that reach several kilometres, imposing large drill rig costs that can be exacerbated by operational drilling problems. Few deep geothermal wells have been drilled, limiting the data available to de-risk these operations. Fortunately, the drilling of hydrocarbon exploration boreholes is a good analogy for geothermal drilling, and digital data from these legacy boreholes are increasingly openly available from National Data Repositories (NDR). We present work from the EU Horizon programme's OptiDrill project (101006964), leveraging these data to develop an unsupervised machine learning method for automatic identification of drilling problems. Specifically, we describe the selection from NDR archives of digital drilling and logging data, the application to these datasets of an Isolation Forest algorithm, and the interpretation of results to identify intervals of anomalous behaviour during drilling. By examining daily drilling reports from the NDR, we demonstrate these anomalous intervals' association with drilling problems and examine their causes. Results could permit avoidance of drilling problems by inspiring remediation measures appropriate to given conditions, or their amelioration through early identification, both reducing the costs of implementing deep geothermal district heating and demonstrating the value to practical geoscience of modern machine learning techniques.