Michael Bonner, formerly of Advanced Energy Industries; and Don Nowakowski, IBM Microelectronics

High-density plasma chemical vapor deposition (HDPCVD) process tools are used to deposit intermetal and postmetal dielectric film layers on IC wafers. HDPCVD tools consist of a single wafer chamber that uses an RF-based, high-density plasma to deposit dielectric films on semiconductor wafers. During deposition, the walls of the process chamber and the process kit hardware become coated with the dielectric film, which must be removed when it reaches a certain predetermined accumulation level. If the dielectric film is not removed completely, it can jeopardize on-wafer film uniformity and thickness, increasing particle levels and, consequently, affecting product yields.

Film removal is accomplished with an in situ RF-plasma-enhanced chamber clean that consists of three steps: a nitrogen fluoride (NF₃)–based high-flow clean, an NF₃-based low-flow clean, and hydrogen passivation. The duration of the cleaning process is based on a deposition- and etch-time calculation that assumes constant deposition and etch rates. However, determining the process endpoint based on that assumption is problematic. Because chamber walls and kit hardware present a variety of surface conditions and geometries, deposition and etch rates are not constant. In addition, etch rates increase throughout a cleaning cycle. Therefore, it does not take twice as long to completely remove 20,000 Å of a given film as it does to remove 10,000 Å of the same film.

Since deposition rates and etch rates are neither constant nor linear, inefficient plasma cleaning often results. Sometimes, the clean runs too long and causes substantial overetching of chamber and kit hardware. Other times, the clean ends prematurely and the film is not removed completely. These inefficiencies lead to costly process tool downtime, undue kit and chamber wear, excess gas consumption, substantial on-wafer particle counts, and low tool throughput, which may reduce product yields.

A further problem with HDPCVD chamber cleans is that gas-flow levels have not been optimized adequately. Typically, cleaning recipes call for a higher gas-flow rate than is necessary for efficient film removal. Excess, unreacted NF₃ is wasted, resulting in unnecessary gas costs. Also, optical spectroscopy, which the semiconductor industry has used for more than a decade to determine chamber-clean endpoints for other types of tools, has not been implemented successfully on HDPCVD tools because of their high gas-flow level during the high-flow cleaning step and low operating pressure during the low-flow cleaning step. Typically, optical spectroscopy does not yield accurate and repeatable endpoints under such conditions.

This article details an HDPCVD chamber clean process optimization effort that addressed these problems by integrating on-line RF metrology technology developed by Advanced Energy Industries (Fort Collins, CO). The article describes process monitoring performed on a production tool at IBM Microelectronics (Essex Junction, VT) for all three chamber clean steps. Results showing reduced gas consumption, reduced processing times, and increased tool throughput are presented.

The RF Metrology Solution

The RF metrology implemented in this project is based on the electrical impedance of the process chamber and the plasma. An RF sensor in the RF path is in contact electrically with the entire chamber and kit hardware, and is highly sensitive to the impedance shifts that occur during chamber cleans, independent of gas flow or pressure levels. This high sensitivity enables accurate and repeatable determination of chamber-clean endpoints. At the beginning of a chamber-clean process, the chamber and kit hardware are at a certain electrical impedance; as the clean progresses and film is removed, the electrical impedance stabilizes. Therefore, by monitoring RF signal changes, the chamber-clean endpoint, or termination, can be easily identified.

The ability to identify when film removal is complete enables users of the metrology to optimize chamber cleans for processing time and gas-flow levels by reducing these parameters in a stepwise manner until the lowest possible levels yielding successful results are determined. The sensitivity and accuracy of the RF sensor also permits optimization of the chamber clean's subsequent hydrogen passivation step. Because none of the three HDPCVD chamber-clean steps have been successfully optimized in the past, the ability to do so represents a significant technology advance.

The RF system used for the optimization effort described here consists of an RF sensor, its corresponding electronics box, and a software package that can be operated on a standard desktop or laptop computer. The RF sensor contains current and voltage transducers and is designed to survive the harsh conditions at the powered electrode without inducing process shifts in key process parameters. The sensor measures the fundamental and harmonic voltages, currents, and phase angles of the process chamber and the plasma.

Figure 1: Schematic showing the in-line installation of the RF sensor used to measure shifts in chamber and plasma impedance.

Typically, the sensor is installed between a CVD tool's impedance-matching network and process chamber. For this project, however, the sensor was installed in-line between the RF power generator and the impedance-matching network, as illustrated in the schematic in Figure 1. The sensor is connected to an electronics box, which is connected to a computer running the RF metrology software. The software enables the viewing and storage of a multitude of RF signals taken from the sensor. Data can be collected at a variety of user-defined rates as fast as 10 milliseconds.

HDPCVD Chamber-Clean Optimization

This project focused on optimizing the chamber-clean recipe of an HDPCVD process tool on the wafer fabrication line for both gas-flow level and processing time. All three chamber cleaning steps were analyzed using RF metrology.

Process-Time Optimization. After an RF metrology system configured for passive monitoring of the chamber cleans was installed on the HDPCVD tool, data from several hundred chamber-clean runs were analyzed. It was determined that the cleans were less than efficient. Both the high- and low-flow cleaning steps resulted in substantial etching. The clean's hydrogen passivation step was also determined to be much longer than necessary.

Figure 2: RF metrology results for a high-flow chamber clean from the passive monitoring phase of the study. The X denotes when film removal was complete.

Figure 2 is an RF metrology plot of one of the high-flow chamber-clean runs. For this cleaning step, the recipe specified a process time of 93 seconds, which was based on the time calculation method typically used for this HDPCVD tool. The X and corresponding drop line at approximately 47 seconds denotes when film removal was actually complete. The RF signal had leveled out at this point and continued to remain stable throughout the remaining 46 seconds of the clean. Therefore, the etching that occurred during those seconds was performing no useful purpose. In other words, 46 seconds of gas flow and tool processing time was being wasted during every cleaning cycle run. Each cycle also subjected the tool to 46 seconds of unnecessary chamber kit wear. Given the high costs of NF₃ gas, process tool time, and chamber kits, the overetching would quickly result in substantial losses.

Figure 3: Results of the overetch experiment. The high-flow clean overetched by ~40% throughout the range of film accumulations tested.

To investigate the universality of these results, an overetch experiment was performed. RF data from chamber cleans of varying durations and film accumulations were analyzed to determine how much overetching would occur based on the current operation of the process tool. Figure 3 presents the results of these tests. It can be seen that the high-flow clean step overetched by approximately 40% throughout the range of film accumulations and process times.

Figure 4: RF metrology results for the low-flow chamber clean that followed the high-flow clean depicted in Figure 2.

On further review of the collected RF metrology data from the chamber-clean runs, it was noted that substantial overetching was also occurring during the low-flow cleaning step. Figure 4 displays the metrology plot from the low-flow step of the chamber clean run that yielded the high-flow clean results shown in Figure 2. This low-flow step was being run from 93 seconds to approximately 141 seconds, a duration that was also determined using the time-calculation method. Again, the X and its corresponding drop line displayed when film removal was actually complete. In this case, the process endpoint occurred at approximately 121 seconds, resulting in 20 seconds of overetching and its associated costs.

Figure 5: RF metrology results for the hydrogen passivation step of the chamber-clean process cycle depicted in Figures 2 and 4.

The review of the collected RF metrology data from the chamber cleans also revealed that a significant amount of processing time was being wasted in the hydrogen passivation step of the chamber clean. Figure 5 displays the metrology plot for the passivation step of the chamber-clean run that yielded the data in Figures 2 and 4. This step was being run from 141 to 280 seconds; the X and its corresponding drop line show that passivation was already complete at 180 seconds. Excess processing in the hydrogen passivation step does not harm the chamber and kit hardware and hydrogen is not nearly as costly as the NF₃ cleaning gas. However, under the cleaning recipe that was being used for the HDPCVD tool, 100 seconds of processing time, per clean, were being lost.

Figure 6: Comparative RF metrology results for a high-flow process clean step using gas flows of 900 and 1000 std cm3/min, respectively.

Gas-Flow Optimization. In the chamber-clean recipe being optimized, the NF₃ flow rate during the high-flow clean step was 1000 std cm³/min. A test was performed to determine whether that flow level could be reduced to 900 std cm³/min. When the resulting RF metrology data were compared with those from a clean using the original recipe, no increase in chamber clean endpoint times was seen. Figure 6 is the RF metrology plot of RF signal traces for 1000 and 900 std cm³/min flow levels. The larger flow-level clean is represented by the red trace, while the smaller one is represented by the black trace. The green drop line denotes the endpoint, which occurred at the same time for both cleans. This means that NF₃ gas consumption can be reduced by 10% without lowering cleaning effectiveness. Because of the high cost of NF₃, such a reduction is an important discovery. It also contributes to the industry's effort to reduce CFC (greenhouse gas) emissions. The possibility of even greater flow-level reductions will be investigated.

After the completion of the HDPCVD chamber-clean optimization investigation, the original chamber-clean recipe was modified by reducing the NF₃ flow level in the high-flow clean step from 1000 to 900 std cm³/min. This new recipe has been used on the production line for several months without affecting particle or yield performance. Figure 7 presents a statistical process control chart of particle performance for the process chamber on which the new recipe was implemented. The process time for the low-flow clean step was reduced by 30%, with no ill effect on particle or product yield performance. As the investigation continues, it is expected that a reduction of the passivation step's process time will be implemented as well.

Figure 7: Statistical process control chart for the production-line HDPCVD chamber studied before and after implementation of a lower gas-flow level in the high-flow cleaning step.

Conclusion

During the HDPCVD chamber-clean optimization investigation described in this article, it was determined that the use of RF metrology can reduce processing time and gas-flow levels. It was found that as much as 2.8 minutes of processing time per chamber clean can be saved and the NF₃ flow level in the high-flow clean step reduced by at least 10%. That reduction has been implemented in production without lowering particle or yield performance. A 30% time reduction in the low-flow clean step has also been implemented, and the investigation into reducing the duration of the high-flow clean and passivation steps is ongoing. The cost savings and increased tool throughput resulting from these changes has been significant. The reduction in NF₃ consumption is also of significance to ongoing industrywide efforts to reduce CFC emissions.

---

Michael Bonner is an independent contractor and a former applications engineer at Advanced Energy Industries (Fort Collins, CO), where he specialized in working with customers to optimize plasma processes using RF metrology. Bonner joined AE when the company acquired Fourth State Technology in 1998. He attended the U.S. military academy at West Point for two years and received a BS in mechanical engineering from Clarkson University in Potsdam, NY. (Bonner can be reached at 845/223-4782 or bonner297@cs.com.)

Don Nowakowski is an engineering technician at IBM Microelectronics (Essex Junction, VT). With 19 years of experience at the company as a technician and engineering technician, he works primarily with CVD equipment and also has been involved with RIE, ion implant, ashing, and ion beam processes. He received an associate's degree, specializing in electronics, from Penn Technical Institute in Pittsburgh. (Nowakowski can be reached at 802/288-3359 or nowakod@us.ibm.com.)