A neural net based, in-line focus/exposure monitor

Two parameters that determine the performance of projection steppers for IC fabrication are the defocus distance and the exposure time. Currently, these settings are optimized by visual examination of a test pattern exposed in a matrix of varying focus and exposure settings. An automated approach promises better consistency and reproducibility at a lower cost. In this work we have automated this calibration task by using digital image processing and neural networks. Digital image processing techniques (such as edge extraction and convolution) are used in pre-processing digitized optical images of specific test patterns printed in a 5×5 matrix of varying focus and exposure settings. The results are used to train a feed-forward neural network to recognize the key aspects of the patterns printed under different stepper settings. The trained network takes a single image of a printed test pattern and estimates the actual focus and exposure settings that were used in exposing the test pattern. It can also suggest possible corrections to the focus and exposure settings in order to ensure optimal operation of the stepper. Results show that this method can identify the optimum settings for the stepper. Also, this procedure can be extended to ensure that the exposure and focus settings are optimal on a run-to-run basis in a production environment