Establishment of MLP models for the three-dimensional transport process in reverse osmosis desalination based on high-throughput computing

Freshwater resources are increasingly scarce around the world. Reverse osmosis (RO) is one of the most advanced desalination technologies, accounting for more than 50% of the desalination market. High permeability reverse osmosis membranes can significantly improve water production efficiency. However, concentration polarization due high flux on membrane surfaces leads to the aggravation of fouling, which greatly hinders the application of high permeability RO membranes. Therefore, developing a more efficient optimal design method based on deep learning methods other than traditional manual iterative design methods for membrane modules can accelerate the commercialization of high permeability RO membranes and has great research value. In this paper, high-dimensional nonlinear transport models in consideration of complex feed channels for seawater RO desalination with a high permeability membrane are established with respect to various working conditions. A three-dimensional (3D) multi-physics computational fluid dynamics (CFD) model with high fidelity is established for the fluid flow and salt transport in the feed channel of the RO membrane module. Using a Latin-like hypercube sampling method, 726 CFD models with respect to different design parameter combinations are generated and calculated on the "Tianhe-2" supercomputer. Theoretically, the maximum computational scale in parallel is over 20000 cores. Then we establish multilayer perceptron (MLP) models incorporating with the CFD simulation results, which are randomly divided into training, validation, and test sets according to the ratio of 650:66:10. The established MLP models using the deep learning method are applied to predict the distribution of 3D local velocity, pressure, and concentration with respect to large-scale parameter combinations within the considered design parameter space, as surrogate transport models for physical models. In order to appraise the accuracy of MLP models, this paper compares the results of 10 working condition models in the test set by calculating structural similarity, root mean squared error (RMSE), and regression coefficient of determination between results using CFD and MLP methods. With calculating RMSE, the prediction accuracy reaches 93.5%, 98.3%, and 95.1%. The calculation efficiency is increased by 1-2 orders of magnitude than that of using the traditional finite element method. To evaluate the accuracy of the results predicted by MLP in terms of macroscopic physical parameters, we calculate the Sherwood number and drag coefficient using the CFD and MLP methods with respect to various working conditions corresponding to the test set. The average prediction errors of the drag coefficient and Sherwood number are 2.68% and 6.23% respectively. Furthermore, the reliability of the MLP models is verified for predicting the 3D local velocity, pressure, and concentration in the reverse osmosis channel under different working conditions. The highly scalable parallel calculation method proposed in this work is suitable for the massive screening and optimal design of multi-type feed spacers. This paper provides a computable model and data support to reveal the complex mechanism of the flow and transport process in the RO membrane module, and has critical significance for the factory-scale application of high permeability RO membranes in the future.