Neural networks (NNs) are used to predict characteristics of GaAlAs layers grown by organometallic chemical vapor deposition (OMCVD). Traditional statistical techniques fail because there are many parameters which control the growth process and relatively few experiments to allow a full description of the effect of changing parameters. A successive approximation technique with NNs was developed which enables the most relevant input parameters to be selected first by a linear NN and then used by a more general NN to accurately predict the layer characteristics. In addition, by training to predict the correction to analytic approximations for the layer characteristics, maximum use is made of prior knowledge about the problem which results in a significant improvement in predictive capability beyond the simple analytic approximations.
Read full abstract