A methodology for selecting process variables for feedback control with application to reactive ion etching

摘要

In many semiconductor manufacturing processes, the important product characteristics cannot be measured in real-time and therefore cannot be directly controlled using real-time feedback control (RFC). However, the benefits of RFC, which include variation reduction through disturbance rejection, may be gained by feedback of process variables closely related to the product characteristics. Reduction of variation in product characteristics will boost process quality, improve yield and facilitate more aggressive product designs. Typically, selection of the process variables used for feedback is based upon qualitative knowledge about the process, which in many cases is limited.;A pair of methodologies has been developed for selecting the best process variables for feedback in order to minimize variation in the product characteristics. These methodologies utilize statistical analysis of data from designed experiments. Three factors contribute to variation in the product variables: the strength of the relationship between the potential feedback variables and the product characteristics, the effect process input adjustments have on the product variables, and sensor noise. The first methodology, Methodology 1, quantifies the first factor as the basis for feedback variable selection. The second methodology, Methodology 2, quantifies all three factors as the basis for feedback variable selection.;In Methodology 1, a set of designed experiments is executed while as many process variables as possible are recorded. The disturbance variables are included as factors in these experiments. Once the data are collected, different combinations of process variables are compared for their ability to predict the product characteristics. A subcategory of regression analysis called subset selection is applicable for this task. A new subset selection approach, labeled 'constraint-limited exhaustive search', is found for this purpose to perform superior to other techniques including principal component regression and partial least squares regression. The results of the subset selection are used as the basis upon which Methodology 1 recommends feedback variables. Methodology 2 uses Methodology 1 to generate a model of the process. This model is used to estimate the variation in the product characteristics for every possible combination of feedback variables. All three factors which contribute to variation are represented in these calculations. These calculations are the basis upon which Methodology 2 recommends feedback variables.;Analytical, simulation and experimental results are presented. The methodologies have been applied to three classes of linear processes using simulated experiments. Both methodologies performed much better than random selection of feedback variables and Methodology 2's performance is near optimal. The advantage of Methodology 1 is that it is much more easily implemented. The methodologies also have been applied to the process of etching a self-aligned gate in a reaction ion etcher. These experiments and the corresponding analysis are presented.