Connectivity Matrix-based Descriptors with Deep Learning for Estimation of Pure Component Properties

Physicochemical properties are fundamental for the design of chemical products and processes. Various approaches including cheminformatic and graph-based methods have been applied for the estimation of pure component properties. In this work, we coupled our connectivity matrix-based molecular structure representation method with a deep learning neural network to develop molecular structure-property relationship models. Molecular structure information is represented by the connectivity matrix, while substructures information is represented by the extracted submatrix. This extraction does not cause any loss of structural information, substructural information can be recovered using the stored matrix. Matrices are transferred into features and characterised based on their eigenvalues. Statistical and mathematical approaches, followed by deep learning neural network models, are developed to correlate the structure-property relationship. Results on the case studies of normal boiling point show that the accuracy of the deep neural network model improved by 9.48% over the previous neural network model.