Estimation of molecular weight distribution of CIS-polyisoprene melts from dielectric loss spectra using neural networks

While the prediction of dielectric loss spectra (DLS) from molecular weight distributions (MWD) is relatively straightforward the inversion is known to be an intrinsically ill-posed problem with high sensitivity to measurement noise. We propose artificial neural networks to solve this problem in two steps: First, the measured DLS is approximated by a special basis function network (BFN), thus reducing the data considerably and inherently smoothing the spectra. Second, a group of simple feedforward networks is employed to estimate the parameters of another BFN. The output of this second BFN is the estimate of the MWD. A simulation demonstrates the performance of the new method.