Optimization Of Preparation Conditions Of Separation Membranes Using Neural Network

Optimization of membrane preparation condition which can enhance membraneperformances and expand application range of membranes is of great significance. Nowadays,it mainly relies on single factor experiment and orthogonal test. That is, membranes undervarious preparation conditions are fabricated and characterized. According to the membraneperformances, preparation conditions are optimized. Obviously, this method has disadvantagesof large experimental data scattering, and the optimal preparation condition is not guaranteed.Therefore, there is an urgent need to develop a universal mathematical model to capture therelationship between preparation conditions and membrane performances. In this study,Flory-Huggins theory was employed to analyse the maximal addition amount of additive, andseveral membranes were fabricated, then image analysis and determination experiment of purewater flux and rejection ratio were used to characterize the membranes. Based on theexperimental data, hybrid models based on backpropagation neural network (BPNN) andgenetic algorithm (GA) were established to optimize the preparation conditions ofpolyetherimide (PEI) ultrafiltration membrane via dry/wet phase inversion. The hybrid modelspredict performances of PEI ultrafiltration membranes and polydimethylsiloxane(PDMS)/ceramic pervaporation composite membranes under various preparation conditions.The hybrid models can contribute to designing the preparation conditions to obtain desiredmembrane performances and avoiding large experimental data scattering in the fabrication ofmembranes.Flory-Huggins theory was employed to investigate the thermodynamic mechanism of thecasting solutions whose polymer is PEI, solvent is N, N-dimethylacetamide (DMAc),nonsolvent is water (H2O), and additives are diethyl ether (DE), polyethylene glycol(PEG400),1-Butanol (BuOH) and1,4-butyrolactone (GBL). The binodal curves and thespinodal lines for H2O/DMAc/PEI, DE/DMAc/PEI, PEG400/DMAc/PEI andBuOH/DMAc/PEI systems were calculated, which reveal the maximal addition amount ofadditive H2O, DE, PEG400and BuOH. According to the thermodynamic property of thecasting solution, preparation conditions of membranes were primarily explored. A statisticalprocedure was developed to measure microstructure parameters of membranes. It disposed thescanning electronic microscope (SEM) images, including gray translation, binarization andnoise reduction of the images, and then maximal pore size, discrete pore size and average poresize were obtained. It was proposed that using discrete pore size and sieving equation tocalculate rejection curve, molecular weight cutoff and molecular size cutoff of the membrane was more accurate that using average pore size, which was proved by the rejection experimentsof bovine serum albumin (BSA) and dextran.Based on the experimental data, BPNN and radial basis function neural network (RBFNN)models were constructed to capture the relationship of five key prepatation conditions (PEIconcentration, additive type and concentration, evaporation time in air and temperature ofcoagulation bath) to the performance of membranes, i.e., pure water flux and BSA rejectionratio. BPNN is easy to be convergent at suboptimal solutions, and there are great deviationsbetween several convergences. RBFNN has dissatisfactory fault tolerance. Therefore, hybridmodels which united perfect local convergent ability of BPNN and ideal global searchcapability of GA were proposed, whose model arctectures were optimized by trial-and-errormethod. Membrane formation mechanisms of various additives are numerous, and theperformance trends were different, but the predictions of the hybrid models were accurate, withmost of deviation in testing data less than10%. The hybrid models can predict membraneperformances under different preparation conditions and hereby indicate H2O/DMAc/PEI/GBLis the best of six casting systems in the study. In addition, the hybrid models can contribute todesigning preparation conditions to obtain higher performances of ultrafiltration membranes(BSA rejection is80-90%and pure water flux is up to1.15-0.95m3m-2h-1) and avoiding largeexperimental data scattering in the fabrication of phase inversion membranes.The hybrid models were used to optimize the preparation conditions of PDMS/ceramiccomposite membrane with the experimental data from the literature to discuss the modeluniversality. The maximal deviations of the training and testing data between the experimentsand the hybrid model predictions were11.01%and7.17%, smaller than those between theexperiments and response surface methodology (RSM) model predictions in the literature. Theaccordances between the the experiments and the hybrid model predictions show that thehybrid models have sufficient accuracy. Connection weight analyses show that PDMSconcentration, crosslinking agent concentration and dip-coating time have great influences onthe performances of pervaporation membranes. In addition, the models predict the preparationconditions to fabricate pervaporation membranes whose permeation fluxes reached10.5and9kg m-2h-1and selectivities were6and7.