| Authors |
Pengju XING, Eric EDELMAN, Carlos VEGA-ORTIZ, Clay JONES, Stephen DEOREO, Mitch STARK, Justin WRIEDT, John MCLENNAN |
| Abstract |
Drilling is a major contributor to the overall cost of geothermal power development. Funded by U.S. Department of Energy (DOE), a research well was drilled in The Geysers Geothermal Field to demonstrate increased drilling performance with polycrystalline diamond compact (PDC) bits. Both PDC and roller cone drill bits were used to drill this well. The objective of this study is to analyze the drilling performance in relation to the geological characteristics of the site. Among other factors, drilling performance is related to the mechanical properties of the rock, including the bulk matrix, lithology interfaces present in situ, and natural fractures or veins. Key challenges encountered during drilling include lost circulation in the mud-drilling section and issues with interfacial severity in the deeper air-drilled section. To fully understand these challenges, it is essential to assess the mechanical properties of the rock formations encountered during drilling. However, a direct measurement of the mechanical properties is difficult. Even for this research well, petrophysical data are limited and there are only sonic and image logs for the shallower mud-drilled section (12.25" hole from approximately 2,400 to 3,300 feet). For the deeper section (8.5" hole, steam reservoir) drilled with air, the only geologic data comes from mud log of the rock cuttings, providing mineralogy and lithology information and indirect assessments of infilled fractures. In this study, machine learning techniques were employed to establish correlations between drilling and geologic data. Firstly, unsupervised machine learning, specifically clustering methods, was applied to identify critical zones associated with drilling issues. By clustering the sonic log data, we were able to identify areas correlated with measured lost circulation, while clustering mineralogical data from the mud logs helped locate zones of interfacial severity (characterized by a sudden drop in weight on bit and an increase in the rate of penetration). Interestingly, not all lithology interfaces from the mud log corresponded to interfacial severity, but the clustering of mineralogy did. This technique is also applied to other offset wells, helping to identify fractures and interfaces within the reservoir. Secondly, supervised machine learning models (e.g. Random Forest) were used to build correlations between drilling data and rock strength. A continuous rock strength profile from the upper section can be inferred from sonic logs, but needs to be calibrated by core testing results. For the deeper section, the lithology and mineralogy from the mud log needs (~3300 ft to 9220 ft) to be interpreted to correlate with the rock mechanical properties. To address this, laboratory testing, including unconfined compressive strength (UCS) tests, triaxial compression tests, and X-ray diffraction (XRD) mineralogy analysis, was conducted using cores from an offset well at The Geysers. These results were used to calibrate the inferred rock strength from the sonic logs, and establish relationships between rock mechanical properties and mineralogy. Consequently, the continuous mechanical properties in the deep section can be inferred from mineralogic data acquired from mud logging. By training machine learning models on these correlations, we can predict both the drilling behavior and formation properties for future wells, as drilling data and mud logs are typically available for each well. With abundant data available for model training, this approach holds significant potential for improving drilling efficiency and predicting subsurface conditions in The Geysers geothermal fields. |