Computational neural networks and the rational design of polymeric materials: the next generation polycarbonates

We present an atomistic approach to computer aided molecular design that incorporates computational neural networks as the tools for determining accurate structure-property relationships. A general computational technique, called PropNet (where Prop represents the type of relaxation, transition, physical, or mechanical property) is developed and applied to the elucidation of transition (glass; Tg, and degradation; Td) and relaxation (Tγ) temperatures, as well as to various other physical and mechanical properties of polymeric materials. TgNet is a general purpose neural network system that rapidly formulates structure-property relationships for the glass transition temperature of amorphous and semicrystalline polymers. Up to 320 different polymers :(Tg ranging from 50-700 K, including tactic and crosslinked polymers) were used to test TgNet for robustness and accuracy. The results demonstrate that TgNet is capable of predicting the glass transition temperature to with 10 K of experimentally reported values. The overall approach can easily be extended to any property for which quality data are available. Individual expert networks: TgNet (glass transition temperature), TdNet (degradation temperature), TγNet (secondary, Tγ relaxation temperatures), RiNet (refractive index), TenNet (tensile strength), ElongNet (maximum % elongation), CompNet (Compressive strength), HardNet (hardness), and IzodNet (Izod notch strength ) were developed and tested. Based on these neural network techniques a number of next-generation polycarbonates for increased impact resistance were rationally designed.