Use of influence diagrams and neural networks in modeling LPCVD

An adaptive learning architecture has been developed for modeling manufacturing processes involving several controlling variables. Experimental results of applying the new architecture to process modeling and recipe synthesis for LPCVD (low-pressure chemical vapor deposition) of undoped polysilicon are described. Control parameters considered are pressure, temperature, gas-flow rate, wafer position, and time. Models for both deposition rate and final mechanical stress in the film have been developed. By using the generalization ability of neural networks in the synthesis algorithm, this architecture can produce new recipes for the process. Two such recipes have been generated for the LPCVD process. One is a zero-stress polysilicon film receipt; the second is a uniform deposition rate receipt based on the use of a nonuniform temperature distribution during deposition