Big data analysis of hollow fiber direct contact membrane distillation (HFDCMD) for simulation-based empirical analysis

Abstract A large number of hollow fiber direct contact membrane distillation (HFDCMD) are simulated using previously developed software, hfdcmd, i.e., a module of an environmental software package (EnPhySoft). Of 11,059,200 cases, 7,453,717 cases are found to be physically meaningful for practical applications. The self-organizing map (SOM) and multiple linear regression (MLR) methods were used to statistically analyze the big data. Using the raw data, physical and dimensionless data sets were prepared with specific formats: the former identifies the most significant parameters, and the latter compares relative importance between input parameters. SOM analysis did not provide transparent dependencies between inputs and/or outputs of HFDCMD: instead, it helped categorize parameters into groups of similar characteristics. Using MLR, we found that macroscopic quantities such as temperature and radii of lumen and shell sides were more influential on the MD performance than microscopic quantities such as pore size and membrane length. A rough (order-of-magnitude) prediction for heat and mass fluxes requires only four key input parameters.