Predicting polymer properties using neural networks

Seeking novel approaches for predicting bulk polymer properties directly from a knowledge of molecular-level properties, values of T_g and tensile modulus for the subject polyimides were calculated based on input of seven molecular descriptors using both partial-least-squares (PLS) multivariate regression and artificial neural networks (ANNs). The residual standard deviation (RSD) between calculated and experimental values of T_g was 17 K from PLS and 22 K from ANN. The corresponding RSD for tensile modulus was 0.15 GPa from PLS and 0.12 GPa from ANN. For both T_g and E, the molecular descriptor with the major contribution to the PLS and ANN models was the `number of rotational degrees of freedom' within the repeat unit.