Optimization of Wavelet-Filtered In-Situ Plasma Etch Data Using Neural Network and Genetic Algorithm

摘要

A new model of plasma etch processes is presented. The model was constructed by applying the backpropagation neural network and genetic algorithm (GA) to wavelet filtered data. During a plasma etching of oxide films, optical emission spectroscopy (OES) data were collected, and the etch responses measured include an etch rate, a selectivity, and a nonuniformity. Discrete and continuous wavelets were applied to filter OES data. Dimensionality of filtered OES data was then reduced by applying the principal component analysis with three variances of 100, 99, and 98%. For each response or each data variance, three types of model were constructed. In other words, a total of 54 models were built for comparison. For the discrete wavelet-filtered data, the optimized model errors are 252A/min, 3.1, 0.51%, identified at 98, 99, and 99% for the etch rate, profile angle, and nonuniformity, respectively. For the continuous wavelet-filtered data, they are 329A/min, 3.83, and 0.31%. Therefore, for the etch rate and selectivity, the discrete wavelet data yielded improved prediction. For the non-uniformity, the continuous wavelet data produced much better prediction than the discrete wavelet data. Compared to earlier models, improved prediction of the proposed model was demonstrated for all but the etch profile angle data.