Enhancing Drilling Fluid Lost-circulation Prediction: Using Model Agnostic and Supervised Machine Learning

Drilling operations are used to exploit subsurface resources for oil and gas production and geothermal energy, resulting in a significant increase in operational costs. As a result, various fluid loss circulation materials (LCMs) have been introduced to mitigate this issue. The ability to predict mud-loss volume before drilling provides engineers with vital information to select optimal LCM characteristics. This study introduces a robust workflow to enhance mud loss prediction using a model-agnostic and Bayesian-optimised Extra Tree (ET) regressor model. Our research draws on a publicly available dataset, which includes more than 2,800 data points from the Marun oil field in Iran. We trained and tested eight different models using this dataset, with the ET model demonstrating the best performance, yielding an MAE of 12.6 bbls/hr and an RMSE of 24.0 bbls/hr. We conducted a permutation feature importance analysis and Shapley value analysis to better understand the relationship between the input features and the target. We then used the selected features to optimise the ET model, resulting in an even better performance, with an MAE of 0.24 bbls/hr and an RMSE of 1.22 bbls/hr. Our proposed approach outperforms some previously reported models in predicting the lost-circulation drilling fluid.