Machine-Learning Workflow for Fracture Geometry Characterization Using High-Resolution Distributed Strain Sensing During Well Production

Abstract Rayleigh Frequency Shift Distributed Strain Sensing (RFS-DSS) has emerged as a crucial technique for characterizing fracture geometry in unconventional reservoirs by measuring strain responses along optical fibers. Several numerical modeling studies have been conducted to relate the RFS-DSS strain data to the fracture geometry and then to estimate the production performance. However, the issue of solution non-uniqueness remains unaddressed. Multiple combinations of fracture properties can potentially generate identical strain responses, introducing significant uncertainty in interpretation. The objective of this paper is to develop an integrated artificial intelligence (AI) assisted workflow to calibrate uncertain parameters and automatically history match field-measured strain responses and then estimate the well production performance. The workflow is summarized as follows: First, the uncertain parameters and objectives are identified for this study. Uncertainty analysis is then conducted to investigate the impact of these parameters on strain change objectives. Initial numerical cases are generated using the Latin Hypercube Sampling (LHS) method. These cases are subsequently simulated using a coupled fluid flow and geomechanics simulator during production. Next, four strain change objectives, peak value, minimum strain, strain area, and strain width, are calculated, and efficient machine learning models are constructed as proxies for the simulator. Different machine learning models are developed for predicting the strain change objectives: Neural Network (NN) models for peak value, minimum strain, and strain area, and XGBoost for strain width. The Markov Chain Monte Carlo (MCMC) algorithm is implemented in conjunction with these proxy models to identify appropriate candidate samples from the posterior distributions. The P50 solution is then derived after rigorous history matching. Our AI-driven workflow demonstrates remarkable computational efficiency, capable of proposing 50,000 samples and generating comprehensive posterior distributions in less than 20 minutes. Validation of the workflow using synthetic cases reveals that the P50 solutions of fracture geometry derived from the posteriors show excellent agreement with the true fracture geometry. The four strain change objectives calculated using the P50 solutions align closely with the true values of the strain change objectives. The results of this study demonstrate the accuracy and robustness of this workflow for predicting strain change objectives and automatically history-matching field-measured strain changes. Furthermore, for field applications, after history matching the measured strain using this workflow to obtain the fracture geometry, we can then estimate production profiles for various completion designs, providing valuable guidance for optimizing completion design and evaluating production performance.