Abnormal pore pressures often lead to drilling instabilities, such as kicks, blowouts, stuck pipes, and lost circulation. Accurate pore pressure prediction is crucial for well planning and drilling to ensure safety and efficiency. Traditionally, pore pressure prediction is manually executed by geomechanics experts. The procedures are usually complicated and time-consuming due to subsurface complexity, and the results are highly dependent on the executor's expertise. In this paper, we presented a machine learning method to train a digital geomechanics model for pore pressure prediction along wells under planning Firstly, we digitalize subsurface geostructure geometries by using seismic-interpreted data and well tops data as input to train a geostructure machine learning model, which classifies and labels the formation and its lithology. Sequentially, we train a digital geomechanics model utilizing a set of machine learning algorithms with existing data, including geology, seismic, drilled well logs, and drilling data etc. The digital geomechanics properties include two main components: rock mechanical properties, and geological burial history-related formation pore pressure and in-situ earth stress properties. The digital rock mechanical properties, including formation lithology, rock elastic modulus, and rock strength et al., are trained as formation material property models describing changing patterns within each formation. The formation pore pressure and earth in-situ stresses, dominated by the burial and diagenetic history of formations, are trained using physics-guided hybrid algorithms that incorporate formation compaction and tectonic parameters regression. Lastly, pore pressure along any planned well trajectories is predicted using this digital geomechanics model, helping to identify drilling risks, optimize safe mud weight, and consequently improve drilling operations. This learnt geomechanics model, as a digital representation of subsurface geomechanics properties in implicit format, is generated by learning algorithms from data and physics laws. As a result, it can self-learn and improve with new well data that emerges during field development.
This machine learning-enabled pore pressure prediction method was tested and validated in an offshore shale oil field in South China Sea, China. The target shale oil reservoir formation is known as an abnormally high-pressure formation, and thus requires a reliable pore pressure prediction in well planning and operations. The prediction in the planning phase usually takes weeks with traditional manual methodology. Applying our machine learning digital model to this field, the initial digital geomechanics model was trained using the first exploration well data and interpreted field geology surfaces data. It was then used to predict pore pressure for the second exploration well. The prediction results were compared to actual downhole pressure tests and confirmed its accuracy. The digital model was then updated with newly acquired well data and used for next planning well prediction. In this case study, the digital model was trained and updated continuously with actual drilling sequence for the first three exploration wells. The prediction of each subsequent well was compared with actual downhole pressure test results, achieving an average accuracy of 96.2%. Also, this digital model was then applied to a horizontal well planned for the next drilling phase, and the digital prediction results were compared with the manual results, achieving an accuracy of 98.5%. Another significant advantage of this model is its high computational efficiency and reduced need for supervision. The digital machine learning method reduced the pore pressure analysis time from weeks to hours. This case study confirms the efficiency and reliability of such digital methods.
