Prediction of sedimentation concentration profiles in inclined suspension systems: A data-driven neural network framework

In the oil industry, drilling processes are essential for crude oil extraction. During drilling, various challenges may arise, requiring the application of different techniques. Vertical wells are sometimes not the most efficient, necessitating the drilling of inclined wells. This inclination increases sedimentation issues, leading to operational challenges. This study presents a comprehensive approach to predicting the sedimentation behavior of suspensions in inclined systems using artificial neural networks (ANNs). The dataset consists of suspension of an aqueous solution containing 20% (v/v) glass microspheres and glycerin. The dynamics of sedimentation were observed at different inclinations (0°, 30°, 45°, and 60°) to understand the effect of orientation on particle settling. Local solid concentration was tracked over time at multiple normalized heights (z/h) using the gamma-ray attenuation method, yielding 9,600 experimental data points. The artificial neural network was trained using time, inclination angle, and normalized height ratio (z/h) as input variables, with solid particle concentration as the model output. The Levenberg–Marquardt training algorithm was employed to optimize the neural network model, enhancing the prediction accuracy by minimizing the mean squared error. K-fold cross-validation is used to thoroughly evaluate the ANN model, which shows excellent accuracy and generalization. From a physical perspective, increasing inclination accelerates clarification in the upper regions of the column and intensifies sediment compaction near the bottom, reflecting the influence of inclined settling mechanisms. The ANN achieved high predictive accuracy R2 > 0.990 with low error magnitudes: MSE 10− 3 to 10− 2, MAE 10− 2 to 10− 1, and MAPE 0.24% to 9.00%, demonstrating robust performance across all tested inclinations (0°, 30°, 45°, 60°). The findings emphasize the ability of neural networks to model complex sedimentation processes, providing a robust tool for analyzing and predicting the behavior of suspensions in diverse conditions. This framework demonstrates that combining gamma-ray attenuation measurements with data-driven modeling provides an effective alternative to traditional analytical approaches and establishes a methodological basis that can be extended in future work to other sedimentation configurations.