Predicting physical and physical-chemical properties of molecular-based materials using computational neural networks

A computational scheme, which utilizes neural networks, was developed to predict properties of molecular-based materials from chemical structures. The method uses a set of simple algorithms to encode the structure and composition of organic molecules directly into numerical vectors, which is used as input for neural networks. Backpropagation type neural networks are then used to correlate these numeric inputs with a set of desired properties. Calculated results for a series of hydrocarbons, fluorohydrocarbons, and crown ethers demonstrate average accuracies of 0.2-8.1% with maximum deviations of 16-20% for a broad range of thermodynamic, physical, and physical-chemical characteristics (heat capacity, enthalpy, heat of evaporation, boiling point, density, refractive index, stability constants, etc.). In addition, a number of physical and mechanical properties were estimated for polymeric materials and compared with regression analysis. Based on the neural network capabilities of formulating accurate quantitative structure-property relationships, a technique called computational synthesis is suggested for performing materials design.