A model was developed to evaluate and predict normal freezing point of organic compounds.This study was based on the group-contribution method and the neural network ensemble technology.The input parameters of neural network were group numbers of organic compounds and isomeric characterization parameters

the output was normal freezing point.The error saturation (ES) condition

which was caused by the use of gradient descent method

would greatly slow down the learning speed of BP algorithm. Having analyzed the causes of the ES condition

an ESP function was proposed to avoid the ES condition from hidden nodes.The simulation result of using ESP method in hidden layers showed that it would reduce the probability of trapping into local minimum in the error surface and accelerate the convergence speed.The ensemble method was used and the neural network ensemble model was established.The nodes in the hidden layer were chosen rationally through simulation

and the cross-validation method was adopted for preventing the BP network from overfitting

which improved the generalization ability of network.According to the ensemble neural network model established in this paper

the estimation result showed that the network had good stability and prediction accuracy and the estimation absolute mean relative error of 207 samples was 8.62%.