A Robust Workflow for Optimizing Drilling/Completion/Frac Design Using Machine Learning and Artificial Intelligence

Abstract

Abstract One of the biggest challenges in drilling/completion/hydraulic fracturing optimization is determining the optimal parameters in the infinite space of possible solutions. Applying a comprehensive parametric study with various geomechanical properties using both a frac simulator and a reservoir simulator is low efficient. This study proposes a workflow for optimizing unconventional reservoir development using machine learning and artificial intelligence (AI) in conjunction with advanced geomechanical modeling. The workflow consists of four steps: in Step1, appropriate acoustic interpretation models are used for geomechanical and in-situ stress characterization. In Step2, unsupervised machine learning optimizes completion designs based on formation anisotropy and heterogeneity along a well. In step3, a training database is built by generating multiple cases based on various simulations guided by a smart sampling algorithm. Proxy models are trained and validated by feeding the training datasets to supervised machine learning algorithms. Lastly, the tested proxy models are run for a multi-parameter sensitivity study for design optimization. The workflow was validated by a Marcellus field case. First, the newly proposed orthorhombic acoustic interpretation model yielded in-situ stress results more consistent with field measurements than the traditional acoustic models. Second, using the C-Means Fuzzy Clustering, the stage and cluster spacings were optimized to overcome the low cluster efficiency issue led by the current geometric completion design. Last, using the newly proposed smart sampling algorithm, a 200-critical-case database was built and fed into the Neural Network algorithm for training proxy models. After running the proxy models in a random-search algorithm, the optimal design parameter values were obtained statistically, leading to the Return-On-Frac-Investment (ROFI) improved by 22-40% from the current base case. The study introduces a robust four-step workflow combining unsupervised and supervised machine learning to examine high-dimensional multivariable drilling/completion/frac designs efficiently. The new workflow enables the evaluation of the statistical significance of the influencing parameters and, most importantly, their interactions, which have often been neglected in the current simulation-based optimization workflow. Moreover, the trained proxy models can be applied to optimize the design of the current wellbore as well as any other future wells drilled in the same basin in a convenient and time-efficient manner.