Application of Predictive Machine-Learning Optimisation Enables Successful Delivery of Highly Challenging Wells

Abstract In this paper a case application will be described of a unique software solution enabling prescriptive optimisation of well delivery, using an industry-proven sophisticated physics-informed machine learning approach that requires no local training, for predictive identification and characterisation of pending well construction risks in a highly challenging, high consequence operating environment. The solution is agnostic to geology, rig, bit/BHA or fluid in use, and requires only surface drilling parameters and trajectory data to deliver advisory for stuck-pipe avoidance, ROP optimisation and vibration monitoring. The Operator is highly experienced in the challenging environment of shallow water Caspian field exploration, development and production, an area typified by extraordinarily variable geopressure and geomechanical regimes leading to well instability and losses/influxes, and often necessitating drilling with exceptional mud weights over 18ppg and resultant well-formation pressure differentials exceeding 6,000psi with associated issues of differential sticking or chance of barite sag; in addition, profile designs required for optimal reservoir placement regularly expose the wells to hole-cleaning risk. A detailed initial forensic study was conducted on a total of 99 days of historical 0.1Hz indexed drilling data covering a range of activities from near-vertical through extended tangent to horizontal sections, both 8½" and 6", with mud weights of 12-18+ppg. It should be highlighted that this data was not used for pre-/re-training of the core machine learning engine of the application in any way; the application was applied ‘as-is’ towards ingestion, processing and analytical review of the dataset provided by the Operator. This empirical offline review indicated a number of examples whereby potential and actual instances of hole-cleaning due to sudden mechanical instability or progressive solids accumulation, overbalance and differential sticking, and increased trend in rotational and/or lateral friction, may have been pre-empted and appropriately actioned often with several hours and/or stands notice. Areas of ROP optimisation were also noted as achievable with subtle modifications in planned drilling parameters without exceeding external operating constraints. The practical learnings from this exercise are then to be taken to complement subsequent real-time application of the same solution, to consume real-time drilling parameters and populate real-time advisory output via industry-standard WITSML exchange for timely live prediction of elevated potential stuck-pipe risk accompanied by nuanced ROP optimisation and vibration monitoring. In addition learnings from ongoing application may form the basis of refined well planning, from more accurate recommendation of optimal drilling parameters by section or formation, through evolution of local best practice into standard operating practices as knowledge-base grows, to maximizing bit-on-bottom time and hence improved prediction of achievable ROP, and therefore AFE and Technical Limit. The application of novel, bespoke, pre-trained machine-learning through a learning cycle of post-run analytics populating subsequent pre-well planning has thereby set an advanced yet practical standard for optimal real-time well delivery.