| In chemistry,traditional experiments are often associated with increased labor costs.The synergy of machine learning and quantum chemistry speeds up this complex and timeconsuming process.To leverage cross-disciplinary knowledge between machine learning and quantum chemistry,a better understanding of the structure-activity relationship is essential for solving chemical problems.Many feature descriptors have been developed in recent years.This paper adopts the descriptor “Bag of Clusters”(BoC)as the main research target.The unsupervised learning algorithms in BoC and applications of BoC in all kinds of scenarios are focused.The main research works in this paper are as follows:(1)Unsupervised learning algorithms are vital for the construction of BoC basis.Three types of basis construction methods based on clustering algorithms are proposed: divisionbased(K-means,Mini Batch K-means),density-based(DBSCAN,Mean Shift),and hierarchy-based(BIRCH).The above cluster(in chemistry sense)finding methods are used to explore the implicit knowledge in the cluster structure data.The results of the different cluster categories are used as the Cluster Expansion basis.Then the molecular feature representation vector is given according to the contribution of clusters in the molecule.(2)Focusing on the performance of BoC,chemical reactions direction prediction is adopted as an assessment tool.The ability of different BoC basis to represent chemical reactions is evaluated.Three typical classification methods are compared: neural network,models with fewer parameters(SVM,SGD),and ensemble learning(XGBoost,Random Forest).Spectroscopic peak positions prediction is adopted as another assessment tool.For1-D spectroscopic data regression,the improved BoC in this paper also competes with classical descriptors.(3)For the huge amount of data gathered previously,knowledge graph exhibition technology and chemical knowledge are combined to handle the reactions network on Ethylene glycol(EG).According to the results from many criteria in chemistry(thermos dynamics,frontier orbital theory),an EG reactions computing platform is developed to show the molecule-reaction-property relationship and reactions graph.This paper has enriched many new basis construction methods in BoC.Successful experience has been accumulated for the practical application of BoC in chemical reactions direction prediction and spectroscopy prediction.Balancing the complexity of the clustering algorithm and expansion space dimension of basis,BoC via MeanShift(BoCms)shows the best performance on chemical information representation.The accuracy of chemical reactions direction prediction achieves more than 97% for BoCms.The prediction means average error in spectroscopic regression has decreased by 7%.Besides,91770 EG reactions with labels are gathered.The research results of this paper combine machine learning with the theoretical calculation of quantum chemistry from a new perspective,which is a comprehensive exploration of BoC based on unsupervised learning,and provides solid academic support for practical application. |
