化学机械抛光的逐次运行控制

A prototype hardware/software system has been developed and applied to the control of single wafer chemicalmechanical polishing (CMP) processes. The control methodology consists of experimental design to build response surface and linearized control models of the process, and the use of feedback control to change recipe parameters (machine settings) on a lot by lot basis. Acceptable regression models for a single wafer polishing tool and process were constructed for average removal rate and nonuniformity which are calculated based on fifilm thickness measurement at nine points on 8 in blanket oxide wafers. For control, an exponentially weighted moving average model adaptation strategy was used, coupled to multivariate recipe generation incorporating user weights on the inputs and outputs, bounds on the input ranges, and discrete quantization in the machine settings. We found that this strategy successfully compensated for substantial drift in the uncontrolled tool’s removal rate. It was also found that the equipment model generated during the experimental design was surprisingly robust; the same model was effective across more than one CMP tool, and over a several month period. We believe that the same methodology is applicable to patterned oxide wafers; work is in progress to demonstrate patterned wafer control, to improve the control, communication, and diagnosis components of the system, and to integrate real-time information into the run by run control of the process.

In Section II, we brieflfly review CMP process and equipment fundamentals, and identify the diffificulties accommodated through run by run process control. Section III presents the overall control system architecture, and describes the gradual mode run by run control strategy. The polynomial response surface and linear control models developed for CMP are discussed in Section IV. In Section V, we present a pair of simulation and fabrication experiments that demonstrate the importance and effectiveness of model adaptive run by run control. Finally, in Section VI we draw conclusions based on these experiments, and highlight areas where additional research and demonstration are needed.

In the CMP process, the wafer is affifixed to a wafer carrier (via back-pressure), and pressed face-down on a rotating platen holding a polishing pad, as illustrated in Fig. 1. A slurry with abrasive material (e.g., silica particles of sizes from 10 nm–200 m) held in suspension is dripped onto the rotating platen during polish. The carrier and platen rotate at variable speeds, typically on the order of 30 rpm. Tools differ in the number of wafers that may be simultaneously polished; single-wafer, dual-wafer, and other multi-headed tools exist.

The process removes material at the surface of the wafer through a combination of mechanical and chemical action. A typical process goal is to achieve “global” planarization (across tens of mm) by preferential removal of “high” material on the wafer. The planarization of dielectric (silicon dioxide) layers between multilevel metallization steps is one common application. Metal planarization is also often performed.

Fig1

In this work, the product characteristics of concern are the removal rate (corresponding to a controlled amount of oxide polished during the step) and the within-wafer uniformity of that removal rate across the wafer. The removal rate is determined by measurement of oxide fifilm thickness before and after polish at each of nine sites on the wafer as shown in Fig. 2, divided by the (fifixed) polish time. The “removal rate” output is the average of the nine sites on a wafer. The “nonuniformity” output parameter is computed for each wafer as the standard deviation of the amount removed over the nine sites on the wafer, divided by the average amount removed over the nine sites, times 100.

The observed drift in CMP processes, and the availability of post-process measurements, motivate the use of a run by run control strategy. A generic semiconductor control system framework has been under development, and is applied to the CMP control problem. Components of the control system design include the following.

The control architecture of Fig. 4 is expanded in Fig. 5 to highlight the control strategy used in this work. Offline experiments are performed on the CMP tool to build empirical (static input-output) models of the process response. An optimal process recipe is selected that satisfifies (or trades off) design goals; this is used as the initial recipe for process control. Lots of 10 wafers each are planarized in the tool, and measurements of oxide fifilm removal rate and nonuniformity are made on wafers #9 and #10. This information is fed to the gradual mode run by run controller, which adapts the process response models. These updated models are then used to generate a new process recipe which: a) achieves the best (weighted) trade-off among the multiple output targets or b) achieves all targets with the smallest (weighted) change in the recipe. The revised recipe is then used for the next lot of wafers.