Estimation of FMI-derived fracture aperture from conventional petrophysical well logs applying ensemble machine learning methods

Studying fracture aperture can yield valuable insights, including detecting high production rate zones, fluid flow and production rate. Conventional techniques are applicable to obtain fracture aperture. However, they are expensive and time-consuming. Innovatively, an integrated, robust, intelligent model is developed to address the challenge of accurately estimating fracture aperture by applying full-bore formation micro imager (FMI) and well-log data from the GHS oilfield (Iran). The model reaps the benefits of the hybrid, ensemble, boosting and tree-based standalone machine learning (ML) algorithms integrated into the optimisation committee machine (CM) and multi-variable linear regression (MVLR) algorithms applying a two-step CM sequence. Six standalone ML models were employed for the initial prediction. Subsequently, four optimisation algorithms were employed within the CM configuration to integrate standalone algorithms, improving the accuracy of fracture aperture predictions by assigning weight coefficients to each algorithm. The genetic algorithm (GA) slightly outperformed the others based on the mean squared error (MSE) and correlation coefficient (R). Utilisation of the CM with GA (CMGA) substantially minimised MSE by 64.48 % (from 0.0020 to 0.0007220) and improved R by 5.68 % (from 0.8971 to 0.9480) compared to the average measurements of standalone models. Further improvement was achieved in the utilisation of MVLR, where all CMs were integrated using the weights derived from the least squares approach. This method unified all CMs into a single structure and enhanced the prediction performance of final fracture aperture estimations with a 1.32 % reduction in MSE and a 0.055 % increase in correlation coefficient compared to the average outcomes of the CMs.