Quantitative Structure-property Relationship Of Ionic Liquids Based On Deep Learning Modeling

Ionic liquid is a kind of green medium and soft material completely composed of anion and anion,which has strong structural design freedom.The combination of cations and anions creates a vast chemical space.Due to the limitation of time and materials,experimental study on the properties of ionic liquids cannot meet the needs of production.Therefore,it has important application value to study the structural-property relationship modeling of ionic liquids to achieve efficient screening of ionic liquids.A self-learning model based on random forest（RF）and convolutional neural network（CNN）is proposed,and the structural-property relationship models of ionic liquid conductivity,acetylcholinesterase inhibition toxicity and sulfur dioxide solubility are constructed by combining regression model.The performance of the models is evaluated.The influence of descriptors on model fitting results is studied by using statistical parameters such as determination coefficient and model error.Based on 1500 experimental conductivity data points of ionic liquids,a stochastic forest ensemble learning model is proposed to quantify the contribution of molecular structure and physical and chemical descriptors to conductivity,and to screen the high contribution descriptors.Then,support vector machine nonlinear regression model（SVR）and multilayer perceptron deep learning regression model（MLP）are combined to model the structural-property relationship of ionic liquid conductivity.RF-MLP model had the highest fitting degree,with a determination coefficient of 0.996,mean square error of 0.79 and mean absolute error of 4.03.The absolute error of RF-MLP model for each data point is within 0.2,which has high accuracy and computational efficiency.The high contribution descriptors screened by random forest model can reduce data redundancy,and can be combined with subsequent regression model to achieve the structure screening of ionic liquids with high conductivity,which has a good application prospect.Based on the acetylcholinesterase inhibition toxicity test data of 266 different ionic liquids,CNN’s unique matrix processing structure is used to model and transfer learning ionic liquid descriptors.Then,feature vectors representing ionic liquids are transferred from the feature extraction layer in the constructed CNN model as input to the subsequent regression model to predict the acetylcholinesterase inhibitory toxicity of ionic liquids.The optimal CNN-MLP model determination coefficient is 0.962,mean square error is 0.015,mean absolute error is 0.103.The relative errors of CNN-MLP model are all within 0.1.The results show that MLP regression model based on convolutional transfer learning can accurately predict the inhibitory toxicity of ionic liquid acetylcholinesterase.Deep learning modeling can realize automatic feature learning and avoid the error caused by artificial feature selection.The solubility of sulfur dioxide in ionic liquid is significantly affected by temperature and pressure.Based on 221 sulfur dioxide solubility data points in ionic liquids,CNN is used to realize self-extraction of ionic liquid descriptors,temperature and pressure.Then,RF,SVR and MLP are combined to model the solubility of sulfur dioxide in ionic liquid.The descriptor learning ability of convolutional neural network is verified by comparing the single model and the forming model.The optimal RF model in the single regression model is determined by 0.946,the mean square error is 0.036,and the mean absolute error is 0.0026.After integrating the CNN transfer learning model,the optimal CNN-RF model has a coefficient of determination of 0.992,a mean square error of 0.0159,and a mean absolute error of 0.0004,which is significantly more accurate than the single regression model.The results show that convolutional neural network can significantly improve the accuracy and robustness of the model,and can be used for computer aided molecular design and screening of ionic liquids.The model provides a new idea of artificial intelligence applied to constitutive relation models of ionic liquids.