Optimizing cementing displacement efficiency using a machine learning ensemble Models: Application of fluent simulation and optimization algorithms

Abstract

Increasing displacement efficiency during cementing has been identified as a crucial measure for enhancing cementing quality. To evaluate displacement efficiency, the correctness of using the Fluent simulation method was verified based on field logging data. On this basis, Fluent numerical simulation results were utilized as the dataset, which formed the foundation for developing prediction models for displacement efficiency. Prediction models were developed using machine learning techniques, employing Random Forest, Extremely Randomized Trees, Artificial Neural Network, and Support Vector Machine algorithms. To ensure seamless integration of these techniques, the Stacking method was applied as a meta-model to combine the most accurate individual models, effectively leveraging their strengths. The prediction accuracy of the ensemble model was analyzed along with the influence weights of various cementing parameters on displacement efficiency. The research further optimized cementing parameters using the Stacking prediction model in conjunction with advanced optimization algorithms, including the L-BFGS, simulated annealing, and gradient descent methods, to achieve optimal displacement efficiency. High computational accuracy was demonstrated by the Random Forest and Extremely Randomized Trees algorithms, with further improvements achieved through the ensemble model combining these two algorithms. The L-BFGS algorithm performed excellently in predicting displacement efficiency in both wide and narrow annular gaps, resulting in improvements of 3.74 % and 5.46 %, respectively, compared to the original method. Furthermore, the deviation between the predicted results and the actual simulated values was controlled within 3.5 %, indicating a high degree of accuracy and reliability. This research not only validates Fluent simulations with field data but also demonstrates the value of machine learning in overcoming measurement challenges. By integrating simulations with prediction and optimization models, it introduces a practical framework that enhances cementing optimization and sets a benchmark for future studies.