Machine Learning Approach Empowers Well Placement in Tight Gas Field

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper IPTC 22188, “Machine-Learning-Empowered Well Placement in a Large Unconventional Tight Gas Field in China,” by Ting Yu, SLB; Xiangzeng Wang, Shaanxi Yanchang Petroleum; and Alexis Carrillat, SLB, et al. The paper has not been peer reviewed. Copyright 2022 International Petroleum Technology Conference. Reproduced by permission. _ A tight gas field produces gas from a large heterogeneous fluvial reservoir with limited well control in margins and appraisal areas and sparse advanced logs. The principal goal of the study described in the complete paper was to provide a novel solution using machine-learning (ML) techniques to predict sandstone distribution and, to some extent, automate the process of optimizing well placement. The presented work flow overcomes low data quality, scaling, and inconsistency and builds the bridge between geoscience and artificial intelligence (AI) software platforms. Introduction The field is an unconventional gas reservoir covering a vast area in the Ordos Basin in China. This field produces gas from a complex and heterogeneous fluvial reservoir with limited well control for prediction of sand-body distribution. Reservoir distribution and quality prognosis carry large uncertainties. The well spacing between exploration and appraisal wells is approximately 3–4 km; well control is sparser in resource-overlaying areas. As such, reservoir prediction using only well data is marked by uncertainties and can result in high geological risk in placing development wells. In addition to well data, 2D seismic lines cover the full field with a spacing of 2×4 km. The described work flow aims to use all geological and geophysical (G&G) information effectively to optimize well placement and minimize risk of dry wells. It can combine G&G software and AI platforms to deepen data mining and analyze correlations between the seismic waveform and sandstone distribution. This is achieved by using advanced ML algorithms to construct a 3D heterogeneity model with multiple variables. It enables quick organization of data for analysis and interpretation and becomes a tool for supporting real-time drilling. Objective The challenge in this fluvial reservoir is to replicate the sand-body-stacking relationships based on sparse and multiscale data integration. With that objective in mind, ML is an effective tool to automate geoscience work flows. In this study, a new work flow was created that takes advantage of ML in subsurface modeling for efficient and computationally inexpensive forecasting. This method incorporates ML in the context of decision support. The main objectives of the study include the following: - Automatically classify seismic attributes into waveforms and establish their correspondence to log motifs - Improve the effectiveness of 2D seismic data integration with well-log data - Generate a reservoir risk map based on seismic and log-data integration - Construct an advanced 3D geological model with multiple variables - Generate a more-objective interpretation by having less human inference in the data analysis - Bridge the gap between G&G and AI software platforms and lift current G&G software limitations This enabled use of advanced decision-tree-based ML algorithms to generate a geological model.