Well-Placement Optimization Workflow Blends Gradient-Free Algorithm, Physics-Informed AI

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 223867, “Hybrid Optimization Framework for Well Placement Using Gradient‑Free Algorithm and Physics‑Informed Artificial Intelligence,” by Kheireddine Redouane, SPE, and Ashkan Jahanbani Ghahfarokhi, SPE, Norwegian University of Science and Technology. The paper has not been peer-reviewed. _ Accurate well placement plays an essential role in increasing field recovery and storage while reducing operational costs. This task is complex, requiring robust solutions that can handle optimization problems efficiently. Despite numerous existing solutions, a need remains for a fast, highly accurate, computer-aided optimization tool. In this paper, an auto-adaptive workflow is described that leverages a complex interplay between machine learning (ML), physics of fluid flow, and a gradient-free algorithm to enhance solving well-placement problems. Introduction The study’s main objective is to develop a novel surrogate-based hybrid intelligent system to handle real well-placement problems in realistic field conditions while overcoming the drawbacks of conventional approaches. The authors’ contribution is the development of a self-adaptive optimization routine (SAOR), a physics-informed artificial intelligence (AI) framework designed to enhance the accuracy and efficiency of the optimization process. This is achieved by directing computational resources to regions of interest identified through surrogate interaction with physics-based simulations by means of the proposed adaptive approach. This approach integrates technically constrained design of experiments (DoE) and a genetic algorithm (GA) optimizer to iteratively enrich the database in the regions where optimization is most critical, ensuring that the surrogate models are continuously refined, leading to faster convergence and higher accuracy. The SAOR in this paper is configured with different types of DoE, two adaptive sampling methods, and two ML models—Gaussian process (GP) and an adaptive neuro-fuzzy inference system (ANFIS). This enables the evaluation of multiple physics-informed AI optimization paradigms that allow the SAOR to explore the “no free lunch” theorem in the context of well placement, demonstrating that the effectiveness of an optimization framework depends on its ability to adapt to the specific characteristics of the problem. To the best of the authors’ knowledge, this is the first study presenting a comparison of direct GA with self-adaptive GP and ANFIS for a well-placement optimization problem. The SAOR demonstrated its capability to design efficient hybrid optimizers tailored to the specific challenges of well placement, achieving high accuracy within reasonable central-processing-unit (CPU) time. Moreover, SAOR ensures that solutions not only are optimal but also adhere to the underlying physics and technical constraints of the reservoir system. Methodology To aid in solving complex well-placement challenges for improved reservoir management, the SAOR has been designed to simultaneously consider the following factors: - Reduction of optimization time by minimizing the number of required simulations, particularly by using DoE and adaptive enrichment techniques - Approximation of the expensive numerical model with a surrogate model that offers faster response times - Dynamic exploration of irregular search spaces, avoiding local optimal results that might lead to suboptimal decisions, and refining the global optimum once it is found