Abstract: An inverse neural model which consists of less input nodes and more output nodes is proposed in this paper.The unique weights of the neural model can be obtained when the number of training samples is greater than the number of input nodes of the model,and they are the pure spectra of components of the mixtures in our experiments.The obtained pure spectra which were extracted from the excitation-emission matrix of anthracene phenanthrene and pyrene multi-sample mixture were compared to the experimental pure spectra.Both obtained and experimental pure spectra were used as calibration in rank annihilation to predict the concentrations of components in the unknown mixtures.The results show that the obtained pure spectra were more suit-able for calibration and the average predictive errors were decreased from 7%,8.5% and 7,8% to 2.3%,5.0%,and 3.5% for anthracene,phenanthrene and pyrene respectively.

Key words: Inverse neural model, Generalized rank annihilation, Excitation-emission spectrum