Taber H. Smith

Run by run control of chemical-mechanical polishing

A prototype hardware/software system has been developed and applied to the control of single wafer chemical-mechanical 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 film 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 tools 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 several months. 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.

Artificial neural network exponentially weighted moving average controller for semiconductor processes

The linear exponentially weighted moving average (EWMA) controller has been shown to improve run-by-run process control for approximately linear processes that are subject to shifts or persistent drifts in the presence of noise. This work addresses the inability of the linear EWMA controller to adequately control processes that are poorly represented by such models. This issue is important to the success of the EWMA controller in semiconductor manufacturing where processes may be poorly approximated with linear process models. We address this issue by outlining an extension of the EWMA controller that utilizes an artificial neural network (ANN) process model. The ANN model is dynamically updated using an EWMA of the biases in the ANN output layer. Recipe generation takes place by optimizing around the dynamic ANN model. We show that this framework improves on the linear EWMA controller for controlling higher order processes. Simulations show that this controller provides stable control for higher order pr...

A self-tuning EWMA controller utilizing artificial neural network function approximation techniques

Recent works have shown that an exponentially weighted moving average (EWMA) controller can be used on semiconductor processes to maintain process targets over extended periods for improved product quality and decreased machine downtime. Proper choice of controller parameters (EWMA weights) is critical to the performance of this system. This work examines how different process factors affect the optimal controller parameters. We show that a function mapping from the disturbance state (magnitude of linear drift and random noise) of a given process to the corresponding optimal EWMA weights can be generated, and an artificial neural network (ANN) trained to learn the mapping. A self-tuning EWMA controller is proposed which dynamically updates its controller parameters by estimating the disturbance state and using the ANN function mapping to provide updates to the controller parameters. The result is an adaptive controller which eliminates the need for an experienced engineer to tune the controller, thereby allowing it to be more easily applied to semiconductor processes.

PATTERN DEPENDENT MODELING FOR CMP OPTIMIZATION AND CONTROL

In previous work, we have formalized the notions of “planarization length” and “planarization response function” as key parameters that characterize a given CMP consumable set and process. Once extracted through experiments using carefully designed characterization mask sets, these parameters can be used to predict polish performance in CMP for arbitrary product layouts. The methodology has proven effective at predicting oxide interlevel dielectric planarization results. In this work, we discuss extensions of layout pattern dependent CMP modeling. These improvements include integrated up and down area polish modeling; this is needed to account for both density dependent effects, and step height limits or step height perturbations on the density model. Second, we discuss applications of the model to process optimization, process control (e.g. feedback compensation of equipment drifts), and shallow trench isolation (STI) polish. Third, we propose a framework for the modeling of pattern dependent effects in copper CMP. The framework includes “removal rate diagrams” which concisely capture dishing height and step height dependencies in dual material polish processes.

A statistical analysis of single and multiple response surface modeling

This work examines the use of single response surface (SRS) and multiple response surface (MRS) techniques for modeling spatial nonuniformity in semiconductor applications. Previous works have suggested that the MRS estimation techniques better measure the nonuniformity due to the underlying spatial function of the process, whereas SRS estimation methods measure the total process nonuniformity (systematic spatial nonuniformity plus random site nonuniformity). This work further highlights this fact in an analytical setting. It is demonstrated that the MRS estimation technique is biased and that this bias can lead to the choice of a nonoptimal process. Experimental data from a chemical-mechanical polishing (CMP) process confirms these observations and demonstrates that careful use of the MRS estimator is required in achieving meaningful results for estimating spatial nonuniformity. Modified versions of each method, which measure spatial nonuniformity alone, as well as versions which measure total nonuniformity, are proposed for the case when one is comparing discrete process settings. Analytical expressions for the expected value and variance of both the SRS and MRS estimators are determined. These are used to compare the efficiency (estimator variance) of these modified estimators. When comparing spatial nonuniformity, it is found that the unbiased MRS estimator is more efficient than the SRS estimator modified to measure spatial nonuniformity. However, it is shown that the MRS estimator, when modified to measure total nonuniformity, is not necessarily more efficient than the SRS method. Finally, the continuous response surface modeling case is considered. It is demonstrated how confidence intervals on the underlying continuous site models lead to a nonuniform bias in the response surface generated by the MRS method. This suggests that care must be taken when using the MRS technique in creating continuous response surfaces of spatial nonuniformity as a function of the process settings.

A study of within-wafer non-uniformity metrics

This work reconsiders within-wafer nonuniformity (WIWNU) metrics for semiconductor processes. Simulations of typical chemical-mechanical polishing (CMP) scenarios are used to demonstrate that these metrics may vary with the pre-process thickness profile, the removal rate characteristics, and processing time. These metrics are compared and contrasted. Some of these metrics are shown to be biased with processing time, while others are shown to be insensitive to improvements in WIWNU. Finally, experimental data is compared with these simulations. It is suggested that multiple metrics may be necessary to determine the actual characteristics of a process.

Optimization of the Chemical Mechanical Polishing Process for Premetal Dielectrics

Since a premetal dielectric (PMD) is used in the first level of interconnects, tight control of the critical dimension of the subsequent first-level contact is essential. The thickness nonuniformity due to PMD chemical mechanical polishing (CMP) can consume most of the depth-of-focus budget for the deep UV lithographic process. In this paper, we first describe a low pressure/high speed CMP process for PMD. Using a polishing pressure of only 2.5 psi, we improve the within die nonuniformity by 20-30%. To meet the throughput requirement, we can achieve a very high polish rate (>5000 A/min) by using high rotational speeds. Second, we describe an integrated noncontact clean to achieve low post-CMP defect counts and metallic contaminations. Cleaning phosphosilicate glass (PSG) is generally more difficult than undoped oxides. Because PSG is softer than an undoped oxide, it usually has more microscratches. Because phosphorous acts as a gettering center for metallic impurities, PSG can have high metallic contamination after post-CMP clean. HF immersion can greatly reduce the metallic contamination, but it enlarges the microscratches, leading to a large number of detectable defects. Mechanisms for the removal of slurry residue, microscratches, and metallic contamination are discussed in this paper.

Wafer Scale Variation of Planarization Length in Chemical Mechanical Polishing

Chemical mechanical polishing (CMP) has become the preferred planarization method for multilevel interconnect technology due to its ability to achieve a high degree of feature level planarity. However, methods are needed to understand and model both wafer level and die level uniformity in interlevel dielectric (oxide) polishing. This paper examines the variation of die level planarity across the wafer and at different process conditions. Substantial dependency of planarization length, a characteristic length which determines die level planarity, on table speed and down pressure is found, varying from 6.2 to 7.8 mm in the experiments considered here. In addition, a dependence of planarization length on die position within the wafer is found, varying by ∼0.5 mm across the wafer resulting in a difference of ∼300 A total indicated range from one die to the next. Some die are impacted even more strongly resulting in much smaller planarization lengths (near 5.0 mm in some cases) due to wafer edge effects. We conclude that accurate modeling and optimization of within-die variation depends on accurate modeling and measurement of not only wafer scale removal rate variation but also wafer scale planarization length variation.

On-line patterned wafer thickness control of chemical-mechanical polishing

We present a gauge study of an on-line metrology system for chemical-mechanical polishing and a 600 wafer run by run (RbR) control experiment enabled by on-line wafer measurement. The variability, reliability, and accuracy of the on-line metrology system are found to be very good. We show that a simple control approach provides a root-mean-squared error of less than 100 A. In contrast, using pilot wafers and sheet film equivalents to control a process results in a 39% decrease in performance, and that using fewer sites may increase variability and lead to an incorrect controlled thickness. We outline how an 8%–80% improvement in throughput, as well as several reductions in cost of ownership, are possible using on-line metrology in conjunction with run by run control.

Bias and variance in single and multiple response surface modeling

This work examines the single response surface (SRS) and multiple response surface (MRS) modeling of wafer uniformity using experimental data from a chemical-mechanical polishing process. It is shown that the SRS estimate of uniformity measures total variation whereas the MRS estimate measures only spatial variation. The experimental results demonstrate that the MRS method is a biased estimator, whereas the SRS method is not. Finally, it is shown that the bias in the MRS estimate can lead one to choose a nonoptimal process recipe.