The Design of Inflow Control Device Completion Using Machine Learning Optimization

Abstract Horizontal wells are often drilled and completed with inflow control devices (ICDs) or autonomous inflow control devices (AICDs) to maximize oil production and prevent early water/gas breakthrough. Conventionally, the design of ICDs/AICDs relies on the 3D dynamic simulations approach, where it requires a substantial number of simulation runs and long processing time. In this paper we present an efficient solution using machine learning (ML) to recommend the optimum ICD/AICD completion configuration to achieve the highest ultimate oil recovery for a horizontal well. Contrasting with traditional methods necessitating hundreds of simulations, the novel ML workflow expeditiously assesses more than 100,000 ICD/AICD configurations using a random sampler. This inclusive approach incorporates both ordinary least squares (OLS) and least absolute shrinkage and selection operator (LASSO) ML methods, efficiently pinpointing the optimal ICD/AICD design. By minimizing the reliance on dynamic simulations, the workflow accelerates the evaluation process. Following ML output, a thorough result revalidation is executed through dynamic simulations, ensuring accuracy. This strategy significantly streamlines the optimization process, offering a robust and efficient means to identify the best ICD/AICD configurations within a shorter time frame. An ML proxy model is created using both OLS and LASSO methods to evaluate the ICD/AICD configuration from a dynamic simulation model and more than 100,000 samples consist of different combinations of ICD/AICD designs from the sampler model to generate the final optimum ICD/AICD configuration with the highest oil production cumulative. This is a huge improvement from the old rule of only looking at a few hundred cases within a certain amount of time. This transformative ability means that working time is cut by three to four times compared to the old ways of doing things, which also leads to a reduction on the computing cost needed for final well setups. In addition to the quantitative gains, the ML-driven workflow stands out because it takes the initiative to reduce nonproductive time (NPT) issues. The faster rate at which optimal designs can be selected protects against possible operational setbacks. The ability to look at more options not only makes it easier to find the best solutions for well deployment, but it also leads to a big rise in oil incremental cases. This shows not only how adopting ML can save you money, but also how important it is for optimizing resources and making operations run more smoothly. The extracted ML outputs were put through a second round of dynamic simulation for revalidation, and it showed very minor differences. The variation in the objective function, which is the oil production cumulative, was less than 0.5%. This method of thorough validation emphasizes how reliable the ML outputs are, suggesting that they should be used to produce the final optimized ICD/AICD completion where the design will be used to deploy in horizontal wells. This study demonstrates that the traditional workflow optimized ICD configurations to achieve a final cumulative oil production of approximately 3.546 MMSTB in over 17 hours. In contrast, the new machine learning-based workflow increased production to around 3.62 MMSTB, cut optimization time to less than 5 hours, and achieved a 2.1% increase in oil production with a 3.4 times reduction in time.