REAL-TIME, IN-SITU ELLIPSOMETRY SOLUTIONS USING ARTIFICIAL NEURAL-NETWORK PREPROCESSING

Development of real time, in-situ monitoring and control of thin film depositions using ellipsometrv requires both rapid data acquisition and rapid processing. We recently developed a numerical solving method fast enough to keep pace with data acquisition. Briefly, the method uses a very fast artificial neural network (ANN) to provide initial estimates to a slower, more accurate variably damped least squares (VDLS) algorithm. The work here addresses a key question raised in the prior work: how the solution workload should best be shared between the ANN and VDLS for fast, accurate solutions. For Ni deposited on BK7 (borosilicate crown glass) substrates, ANN accuracy of 10% in d1 and d2, 0.1 in n1 and 0.1 in k1 resulted in solutions generally under 10 iterations. Training of a network using 2000 data over 2000 presentations was sufficient to achieve this accuracy. Iteration contour plots of VDLS performance combined with ANN target plots provided the necessary information to determine the accuracy values required for proper operation with Ni on BK7 data.