AI-Enhanced Multiphysics Imaging Advances Perforation-Erosion Analysis

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 223570, “AI-Enhanced Multiphysics Imaging for Perforation-Erosion Analysis: A Quantum Shift in the Accuracy and Efficiency of Unconventional Well Optimization,” by Glyn Roberts, SPE, Souvick Saha, SPE, and Shane Moreno, SPE, EV Offshore, et al. The paper has not been peer-reviewed. _ The complete paper demonstrates how the integration of multiphysics downhole imaging with machine-learning (ML) techniques represents a major advance in perforation erosion analysis. The authors describe a novel approach that improves measurement accuracy, consistency, and turnaround time, illustrating how this capability benefits the fields of completion design and optimization of hydraulic fracturing and enables gains in productivity and cost reduction in unconventional-wells development. Application of ML to Downhole Camera Images Before the application of ML models, for a typical well, the analysis process for approximately 500 perforations would often take 3–4 weeks. As the industry has developed and drilled longer lateral wells, it is common to have an average of approximately 1,000 perforations per well. Without the application of ML, it would require 6–7 weeks to fully capture images; dimension perforations; and compute erosion by perforation, cluster, and stage before reporting results. With the application of ML methods, end users of the data now obtain a clearer perception of perforation erosion, an increasingly sophisticated understanding of proppant placement, and an overall assessment of stimulation effectiveness within 2–3 weeks of completing the camera survey. Moreover, because such information helps to improve stimulation design for upcoming wells, expected well productivity is accelerated. Initial Application of an ML Method. A training data set of 35,000 perforation images was used initially to address one of the most time-consuming manual processes. The objective was to more efficiently detect perforations in an image while differentiating them from similar features, such as debris or casing marks that might be mistaken for perforations, and to define their perimeters to provide measurements. To be useful, this task needed to function effectively in challenging downhole conditions of poor fluid clarity, suspended solids, or fluid-phase separation. The data was divided into an approximately 85% (30,000) training set and an approximately 15% (5,000) validation set to ensure effective model evaluation. The ML model was developed using an advanced object-detection framework which, at the time of implementation, had state-of-the-art capabilities. It uses region-based convolutional neural networks and advanced algorithmic tools to analyze and process objects within images. It can perform multiple computer-vision tasks, which are useful in the workflow that was being addressed, including object detection (identifying and localizing objects with bounding boxes) and semantic segmentation (assigning class labels to each pixel for accurate object delineation—in this workflow, the process of defining the perimeter of the perforation). This model was applied specifically to detect and segment perforations in still images rather than in video sequences. Fig. 1 shows a side-by-side comparison of a perforation perimeter identified and then outlined by an analyst (left) and the perimeter detected and segmented by the model (right).