Pipelines are critical infrastructures in industries such as water distribution, mining, energy, ... more Pipelines are critical infrastructures in industries such as water distribution, mining, energy, and telecommunications, yet leaks, blockages, and environmental stresses often compromise their performance. To address these challenges, this study develops a predictive model of flow in a pipe using a hybrid approach that combines numerical simulations in ANSYS with machine learning techniques in MATLAB. One hundred one simulations were performed to generate datasets with input parameters including inlet diameters, velocities, temperatures, and outlet conditions. The output parameters of interest were turbulent (eddy) viscosity, maximum pressure, and maximum flow velocity. These datasets were used to train neural networks with three algorithms: Levenberg-Marquardt, Bayesian Regularization, and Scaled Conjugate Gradient. Results show that Levenberg-Marquardt and Bayesian Regularization achieved near-perfect correlations (R > 0.99) for pressure and velocity prediction, while Scaled Conjugate Gradient provided moderate accuracy. Bayesian Regularization outperformed the other methods with an overall correlation (R ≈ 0.973) for turbulent viscosity, highlighting its robustness in handling complex nonlinear relationships. The findings demonstrate the effectiveness of integrating computational fluid dynamics and machine learning for predictive modeling, offering a pathway toward improved pipeline performance, fault detection, and preventive maintenance.
Uploads
Papers by Jordi Malamba