Application of recurrent neural networks to model the defluoridation process of hydroxyapatite synthesized by simple methods

Highly fluoridated groundwater can pose a serious threat to human health, and in this study, hydroxyapatite (HAP) for defluoridation was readily produced by the wet precipitation method and the characterization confirmed that the adsorbent synthesized was the same as analytical grade HAP in either elemental composition or crystal structure. HAP exhibited a remarkable adsorption capacity for fluoride and reduced the fluoride content of solutions with initial concentrations of 5 mg/L and 10 mg/L to less than 1.5 mg/L when dosed at 1.5 g/L and 4 g/L respectively. The fluoride uptake behaviour of HAP followed the pseudo-second-order model and the Freundlich model and a maximum adsorption capacity of 20.8756 mg/g was achieved at 283.15 K. Both the negative Delta G degrees and the negative Delta H degrees suggested that the adsorption of fluoride on HAP was a spontaneous and exothermic process. In addition, the adsorption mechanism can be attributed to the ligand exchange and internalization. Based on recurrent neural networks (RNN), the Long Short-term Memory Network (LSTM) and the Gated Recurrent Unit Network (GRU) which took into account 11 variables, were created for forecasting the F- removal efficiencies of HAP under different conditions. It can be obtained from the training results (training set R2 > 0.995 and test set R2 > 0.991) and the subsequent validation trials (error rate = 0) that the models possessed excellent forecasting capabilities and robustness. The HAP synthesized by inexpensive materials and simple methods is a feasible adsorbent candidate for removing excessing F- from groundwater, and the pre-diction models developed through RNN reveals high application prospects in the field of predicting the effec-tiveness of defluoridation.