No abstract

This paper presents results and commentary on using a cascade neural network and a policy-iteration optimization routine to provide suggested process setpoints for recovery from long-term machine drift in a LAM 4520 6-in dielectric etcher. Traditional plasma etch variables such as pressures, gas flows, temperatures, RF power, etc, are combined with a generalized representation of the time dependent effects of maintenance events to predict film etch rates, uniformity, and selectivity, A cascade neural-network model is developed using 15 months of data divided into train, test, and validation sets. The neural model both fits the validation data well and captures the nonuniformity in the in-control region of the machine. Two control algorithms use this model in a predictive configuration to identify input state changes, including maintenance events, to bring an out-of-control situation back into control. The overall goal of the optimization is to reduce equipment downtime and decrease cost of ownership of the tool by speeding up response time and extending the lifetime of consumable parts. The optimization routines were tested on 11 out-of-control situations and successfully suggested reasonable low-cost solutions to each for bringing the system back into control.

This paper details the use of neural network technologies in the characterization of bit fail patterns occurring on SRAM chips as an alternative to the more traditional rulebased or knowledge-based approach to fail pattern occurrence and classification analysis. The results of bit fail pattern count analyses are used both for fault analysis post-processing and manufacturing yield improvement methodologies. The move toward neural network implementation comes in response to prohibitively long processing times required for implementation of rule-based algorithms on more complex devices and the added flexibility of a neural network to learn new fail types in a more adaptive mode. An unsupervised approach to fail pattern identification was implemented on a 128 K SRAM chip using a two-layer Kohonen Self Organizing Map for identification and concurrence of bit fail pattern categories within SRAM chips. A second network utilized a multilayer perceptron (MLP) architecture with backpropagation of error for prediction of the number of occurrences per bitmap of each of the 34 previously identified shape types. The MLP used the output of a SOM as its input vector to assist in the feature extraction by shape type. Both trained networks out-performed existing rule-based algorithms both in ability to identify bit fail pattern types, frequency counts, and speed of processing.

This paper details the development of two pattern recognition approaches-modified target factor analysis (TFA) and artificial neural network analysis-applied to Auger electron spectra for precise discrimination of thin-film TIN, compound composition. Use of Auger electron spectroscopy for the analysis of titanium nitride films offers advantages over other spectroscopic methods in its ability to analyze either blanket-deposited films or submicron features on patterned wafers and its speed of sample preparation and spectral processing. However, severe overlap of the characteristic Ti LMM and N KLL transitions between 380 and 390 eV prohibit direct stoichiometry measurement. Modified TFA and multilayer perceptron (MLP) neural network approaches are applied to the task of discriminating TIN, stoichiometries to within 2% of a nominal 1 : 1 composition. The modified TFA procedure succeeded in accurate prediction of AES spectra TiNz stoichiometries in five of six sample groupings. The MLP neural network approach accurately predicted stoichiometries for all samples with reduced sample variance over standard analytical methods. This work combined AES and neural network technologies to improve significantly the precision of the stoichiometric composition recognition capability for TIN, compounds. INTRODUCTIONThis paper details work performed in the development of pattern recognition procedures applied to the precise discrimination of chemical stoichiometrics of thin-film titanium nitride (TiN,) compounds via two methods of mathematical decomposition of Auger electron spectra : multivariate linear target factor and artificial neural network analyses. Auger electron spectroscopy (AES) offers the advantage over other spectroscopic methods such as x-ray photoelectron spectroscopy (XPS) and Rutherford backscattering spectrometry (RBS) in the analysis of titanium nitride films in its ability to analyze either blank-deposited films or submicron features on patterned wafers. In addition, the sample preparation and spectral data processing time of Auger electron spectra can be much less than other spectroscopic methods.' The precision of AES is anticipated to be sufficient to enable discrimination of stoichiometries to within 0.5 at.% concentration differences.The principal difftculty arising in the use of AES for TiN, stoichiometry discrimination is the appearance of major overlap of the characteristic Ti LMM and N KLL transitions between 380 and 390 eV. Because the individual spectra cannot be measured directly, pattern recognition techniques for discrimination of atomic concentrations of Ti and N are required.To date, published AES pattern recognition studies2v3 have had success in compound concentration discrimination by utilizing targeted factor analysis (TFA) methods developed by Malinowski and H~w e r y .~ Stickel, Watson and Diebold's3 findings show Ti and N peak separation and accurate relative concentration

This paper describes a generic dynamic control system designed for use in semiconductor fabrication process control. The controller is designed for any batch silicon wafer process that is run on equipment having a high number of variables that are under operator control. These controlled variables include both equipment state variables such as power, temperature, etc. and the repair, replacement, or maintenance of equipment parts, which cause parameter drift of the machine over time. The controller consists of three principal components: 1) an automatically updating database, 2) a neural-network prediction model for the prediction of process quality based on both equipment state variables and parts usage, and 3) an optimization algorithm designed to determine the optimal change of controllable inputs that yield a reduced operation cost, in-control solution. The optimizer suggests a set of least cost and least effort alternatives for the equipment engineer or operator. The controller is a PC-driven software solution that resides outside the equipment and does not mandate implementation of recommendations in order to function correctly. The neural model base continues to learn and improve over time. An example of the dynamic process control tool performance is presented retrospectively for a plasma etch system. In this study, the neural networks exhibited overall accuracy to within 20% of the observed values of .986, .938, and .87 for the output quality variables of etch rate, standard deviation, and selectivity, respectively, based on a total sample size of 148 records. The control unit was able to accurately detect the need for parts replacements and wet clean operations in 34 of 40 operations. The controller suggested chamber state variable changes which either improved performance of the output quality variables or adjusted the input variable to a lower cost level without impairment of output quality.