Neural network modeling of variable frequency microwave curing

Abstract

A new thermal processing technique, known as variable frequency microwave (VFM) curing, can perform the same processing as traditional thermal curing in minutes without compromising the intrinsic material properties. To verify that VFM curing results in comparable material properties, several experiments are performed with samples of benzocyclobutene (BCB) cured on silicon wafers. Curing is performed in a Lambda Technologies MircoCure/spl trade/ 2100 system, as well as a conventional thermal furnace. The BCB samples are heated to an appropriate temperature and held at that temperature for a specific amount of time for both processing methods. Through-plane and in-plane indices of refraction are subsequently measured using a Metricon prism coupler. Backpropagation neural networks are then implemented to model the VFM curing and the conventional curing processes. Inputs to these neural networks are temperature and time-at-temperature, and the outputs are the in-plane and through-plane indices of refraction of the cured BCB samples. The indices of refraction are then used as metrics to determine the extent of cure of the BCB. To validate the neural network models, the root-mean-square (RMS) error is used as a performance metric. RMS errors on the order of 1e-2 are achieved when comparing the outputs of the neural network models with test data.