Real-Time Physiconeural Solutions for MOCVD

Abstract

This paper presents an integrated physiconeural network approach for the modeling and optimization of a vertical MOCVD reactor. The basic concept is to utilize the solutions obtained from a physical model to build an accurate neural network (NN) model The resulting model has the attractive features of self-adaptiveness and speed of prediction and is an ideal starting tool for process optimization and control. Following this approach, a first-principles physical model for the reactor was solved numerically using the Fluid Dynamics Analysis Package (FIDAP). This transient model included property variation and thermodiffusion effects. Using software developed in house, neural networks were then trained using FIDAP simulations for combinations of process parameters determined by the statistical Design of Experiments (DOE) methodology. The outputs were the average and local deposition rates. It is shown that the trained NN model predicts the behavior of the reactor accurately. Optimum process conditions to obtain a uniform thickness of the deposited film were determined and tested using the physical model. The results demonstrate the power and robustness of NNs for obtaining fast responses to changing input conditions. A procedure for developing equipment models based on physiconeural network models is also described.