Neural Network Techniques Applied to Predict Flammability Limits of Organic Compounds

摘要

The application of the neural networks to correlate lower and upper flammability limits with chemical structures or basic physical properties for ISO organic compounds of diverse structures and functionalities was investigated. The prediction results based on a three-layer neural network with a back-propagation algorithm were compared with those obtained by multiple regression analysis. The comparison showed that the neural networks gave better versatility (range of applicability) and reliability. The predictive ability of the methods developed were tested for a new set of 50 compounds not included in the training data set and good agreement with observed lower and upper flammability limits was confirmed. The required information for the calculation were six easily available properties: standard enthalpy of combustion, molecular weight, critical temperature, critical pressure, oxygen balance, and diffusion coefficient in air. By using these calculable properties from the molecular structure, the method could be used to predict flammability limits of new compounds containing C, H, O, N, S, Cl, F and Br atoms.