Title: Rheology data-driven machine learning models for phosphate slurry pipeline in Morocco
Morocco has adopted phosphate transportation through hydraulic pipelines from Khouribga to Jorf Lasfar. This allows to increase the production and reduce the transportation cost. Given the vital importance of phosphate in global food security and regarding the huge amount of phosphate rock reserves in Morocco, it is detrimental to assess the reliability, to optimize and to increase its transportation in a safe manner. Usually, hydraulic transportation of such fluids is fully quantified with a full characterization of its rheology related to its non-Newtonian behavior. The rheology allows to know the viscous and the elastic properties of a fluid exhibiting viscoelastic properties. In the case of water-phosphate slurry this behavior is not well-documented and classical constitutive laws for the rheology are of limited use, because of the high variability of different physico-chemical components of the slurry. In the present work we propose a deep learning approach that can accurately and efficiently quantify the water-phosphate slurry rheology to these components. To train and validate the deep learning approach a large dataset of measurements is generated from both the head station at Khouribga and the terminal station at Jorf Lasfar. The trained network is then tested for both forward and inverse problems where the produced solutions are compared to their experimental counterparts. Once the proposed network is trained it is possible to establish constitutive laws for the phosphate rheology. Results presented in this study demonstrate that the bagging allows to reduce the validation error of the model by up to two orders of magnitude. Thus, it considerably reduces the variability on the estimation of hyperparameters in the model. Moreover, the sensitivity analysis shows that the variability on the elasticity coefficient is mainly due to the variability of the slurry density and the solid rate. Viscosity on the other side is not affected by the heterogeneity of the granulation distribution.