Development of machine learning and stepwise mechanistic models for performance prediction of direct contact membrane distillation module- A comparative study

Abstract

The development of accurate and fast modeling tools to predict the performance of direct contact membrane distillation (DCMD) modules can result in their performance improvement. Conventional models ignoring the variations of operational parameters along the membrane's length may cause inaccuracy. Moreover, models considering these variations are often complex and computationally expensive. To propose an accurate and fast modeling tool, the possibility of using machine learning models for the performance prediction of the DCMD module has been studied for the first time in this study. The robustness of three machine learning models (ANN, SVR, and RF) has been thoroughly compared with the stepwise mechanistic modeling approach in terms of models’ accuracy, trend predictability, and computational time. The results show that ANN and SVR models exhibit an enhanced performance over the mechanistic model, possessing a MAPEtest of 3.46% and 4.78% as compared to the mechanistic model with a MAPEtest of 7.31%. Further, compared to the mechanistic model, the machine learning models have the privilege of simplicity, enhanced accuracy, and significantly lowered computational time. The feature importance analysis also showed that the feed flow temperature is the most influencing parameter on permeate flux in the DCMD system.