Machine Learning-Based Estimation of Solids Content in Drilling Fluids Through Utilization of a Comprehensive Mud-Report Database

Characterization and optimization of total solids content of drilling fluids is critical for the efficiency and success of drilling operations. Traditional solids content analysis methods, such as retort analysis, require substantial human intervention and time, which can lead to inaccuracies, time-management issues, and increased operational risks. To address these issues and complement automated rheological property measurements during drilling, this study aims to develop and validate a machine learning-based framework for estimating solids content in drilling fluids from readily available rheological parameters. A comprehensive data set was compiled from more than 1600 laboratory reports of drilling fluid analyses across 130 oil wells globally. Text mining packages were used to convert laboratory reports into a numeric data set which consists of drilling fluid properties. Pre-processing steps were taken to clean, filter and reduce the dimensionality of the data. Unrealistic measurements, and missing data were filtered out and several feature-selection approaches such as least absolute shrinkage and selection operator (LASSO) regression, permutation feature importance, and correlation coefficient matrix were employed to remove redundant variables from the data set. Cleaned and dimensionality-optimized data set was then used for machine-learning model development. Due to their ability to capture non-linear relationships among multiple variables, artificial neural networks employing the resilient backpropagation algorithm were selected to fit regression models for estimating total solids content. Various configurations were tested, including different numbers of layers and neurons, as well as alternative data partitioning schemes. Considering reasonable values of rheological properties based on domain knowledge, filtering was applied to remove outliers which indicate experimental error. As a result of combining the observations from different feature importance analyses and considering physical relationships between rheological variables, redundant features as well as reports with many missing values were removed from the data set. Five features, including mud weight, chloride content, plastic viscosity, oil-water ratio and API fluid loss, were selected to carry over into the modeling stage. As a result of experimenting with 120 different neural network configurations, a two-hidden layer neural network with 40-30 neurons in each hidden layer was selected. The model achieved an R² value of 0.96 for the training set and 0.89 for the testing set. The root-mean-square error (RMSE) was 0.87% for training and 1.27% for testing, both of which were considered acceptable levels of error. An analysis of the network weights indicated that mud weight (MW) and oil-water ratio (OWR) were the most influential features, contributing more significantly to the model’s predictions compared to other input variables.