Optimization of neural network structure and learning parameters using genetic algorithms

Neural network models of semiconductor manufacturing processes offer advantages in accuracy and generalization over traditional methods. However, model development is complicated by the fact that backpropagation neural networks contain several adjustable parameters whose optimal values are initially unknown. These include learning rate, momentum, training tolerance, and the number of hidden layer neurons. This paper investigates the use of genetic algorithms (GAs) to determine the optimal neural network parameters for modeling plasma-enhanced chemical vapor deposition (PECVD) of silicon dioxide films. To find an optimal parameter set for the PECVD models, a performance matrix is defined and used in the GA objective function. This index accounts for both prediction error as well as training error, with a higher emphasis on reducing prediction error. Results of the genetic search are compared with a similar search using the simplex algorithm. The GA search performed approximately 10% better in reducing training error and 66% better in reducing prediction error.