Engaged in the global trend towards more energy-efficient and sustainable technologies, our research team has developed a falling film distillation apparatus with an innovative heat supply through a two-phase closed thermosyphon. In order to evaluate the performance of this energy-intensified distillation process, a supervised machine learning (ML) predictive model based on artificial neural networks is implemented for the separation of the ethanol–water binary mixture. The feed temperature, the evaporator temperature, and the feed flow rate are the three input variables of the model, whereas the ethanol mass fraction in the distillate, the distillate mass flow rate, the recovery factor, and the separation factor are the four performance indicators evaluated. The feed-forward ML has been trained, tested, and validated using a total dataset of 64 experimental runs carried out in the pilot-scale unit, covering a wide operating range of the input variables. Despite the high non-linearity, the ML approach was capable of modeling this new process accurately. The optimal topology of the ML model was achieved with a network arrangement of 10 neurons within 1 hidden layer (3:10:4), with a correlation coefficient (R) greater than 0.95 for all data. The predictive abilities of the ML model were harnessed to investigate the individual and synergistic interaction effects of the operating variables by plotting generalization graphs. Finally, the optimal operating conditions were evaluated by the genetic algorithm (GA) technique, being the feed temperature of 90.6 °C, the evaporator temperature of 109.6 °C, and the feed flow rate of 26.3 L/h, the process operating values that led the maximum of the four performance indicators, simultaneously. Under these operating conditions, a distillate mass flow rate of 4.9 kg/h, with 50.6%wt ethanol-enriched, a recovery factor of 84.9%, and a separation factor of 57.4 have been achieved, showing the high-performance feasibility of the pilot-scale unit.
Read full abstract