Construction Of Deep Learning Model And Prediction Of The Performance Of Anion Exchange Membranes

As a powerful technology,machine learning algorithms can greatly promote the development of new materials.Among these algorithms,artificial neural network(ANN)is a kind of model used to establish correlations between qualitative and quantitative parameters.As the data become increasingly huge and complex,the requirements for the model is increased.Recently,many attentions have been paid to the deep neural network(DNN)with the multiple hidden layer structure.At present,the application of DNN in chemistry is still in the preliminary stage.However,it has attracted much attention due to its excellent learning ability.Alkaline anion exchange membrane fuel cells(AAEMFCs)is an efficient and environmentally friendly energy conversion device,which has been considered to be one of the next generation fuel cell systems.Anion exchange membranes(AEMs)are one of the most important components of the AAEMFCs.Great efforts have been made to develop AEMs with superior electrochemical properties.However,the traditional methods of developing membrane materials usually need the experienced researchers to do a large number of experiments by spending a lot of time with high costs.Therefore,it is urgent to construct the DNN model to investigate the correlation between the chemical structure and the properties of the AEMs.The model could predict the performance of AEMs to guide and simplify the experimental studies.In this work,poly(2,6-dimethyl-1,4-phenylene oxide)(PPO)was chosen as the matrix material.A large amount of data was collected from literatures for construction of the model.The thesis includes the following contents:(1)The water uptake prediction model was built based on the collected water uptake data.The influence of IEC,temperature and chemical structures on the water uptake of the AEMS was investigated.The importance ranking between different descriptors and water uptake was determined.The characteristics of 23,which affect water uptake,were screened.The parameters affecting DNN network structure were also investigated,including training set proportion,number of neurons in hidden layer and learning rate.Moreover,a model of multi-linear regression(MLR)was constructed for comparison.The optimized network parameters were used to predict the water uptake of the AEMs based on PPO.The results showed that the Pearson correlation coefficient(Rp)of the fitting curve formed by the experimental value and the predicted value reached 0.944,and the root mean square error(RMSE)reached 0.055.(2)The conductivity prediction model is constructed based on the collected conductivity data.The importance of different characteristics determining the ion conductivity was ranked.The influence of the number of input characteristics was investigated.It is found that DNN model has excellent performance when selecting the top important characteristics of 25.The MLR model was also constructed for comparison.Adjusting the structural parameters of the two models,the trained model can predict the ion conductivity of the AEMs.The results showed that when the DNN is used as the model prediction,the Rp of the fitting curve formed by the experimental value and the predicted value reaches 0.774,which is 0.2 higher than that of MLR model,The RMSE of the DNN model is 0.021.(3)The multitask DNN model was constructed by using the data of conductivity and water uptake.This model could be used to predict both the water uptake and ionic conductivity of the AEMs once adjusting the parameters.The descriptors of DNN model are chemical structure,temperature and IEC.It is found that the DNN model with the optimized network parameters exhibited the same performance as the single-task DNN model.As respect to the water uptake prediction,the Rp of 0.838 and RMSE of 0.021 of the fitting curve formed by the experimental value and the predicted value are achieved,respectively.While the Rp of 0.567 and RMSE of 0.029 are reached to the conductivity prediction,respectively.