Defect/Yield Analysis and Metrology

Measuring and characterizing opaque multilayer metal film stacks on product wafers

George J. Collins, Rudolph Technologies

A nondestructive metrology tool that provides data on thickness and other film properties has applications in yield management as well as process monitoring.

The need to improve yield and avoid process errors has been a part of semiconductor manufacturing since the first integrated circuit was made. According to Jack Kilby, a coinventor of the IC, in the early 1960s a respectable yield on a single transistor was 10 or 20%.¹ Today, wafers containing hundreds of ICs, each having 20 million transistors, can be manufactured with yields exceeding 90%.² This remarkable improvement in yield was made possible by concurrent advances in both process technology and metrology. In transparent-film metrology, for example, tremendous progress has been made since the days when process engineers estimated single-layer film thickness by comparing the color of the deposited film with color charts. Today, ellipsometers routinely support semiconductor production, giving process engineers fast, accurate, and repeatable thickness measurements for each individual layer in transparent multilayer film stacks on product wafers. However, there was no comparable metrology system for the analysis of opaque multilayer metal (MLM) film stacks (as depicted in Figure 1) until 1997, when Rudolph Technologies (Flanders, NJ) introduced picosecond ultrasonic laser sonar (PULSE).³

Figure 1: TEM image of an interconnect structure cross section showing MLM film stacks at levels metal-1 through metal-4.

In addition to providing thickness data for each metal layer in film stacks with as many as five layers, the technique provides information on density, roughness, and adhesion strength for both top and buried layers. Tools based on this technology have already been accepted by chip manufacturers for MLM process control applications. In addition, Rudolph Technologies has partnered with several semiconductor fabs to study the use of such tools to detect metal-deposition processing problems. The information the tools provide enables engineers to understand the cause of these problems and to take cost-effective corrective actions quickly. After a description of how the technology is used in process monitoring, this article focuses on its potential usefulness in yield management applications. The results of several studies are presented and their implications are discussed.

In-Line MLM Film Stack Process Monitoring

Described in more detail in the accompanying sidebar, the new laser sonar technology uses a very brief (100-femtosecond) laser flash to induce an ultrahigh-frequency sound wave that travels from the surface of the top metal layer down through the film stack. As the sound wave crosses the interfaces between each film layer, some of the sound reflects back to the top surface as a series of echoes. When each echo reaches the top surface, it changes surface reflectivity, allowing its arrival to be detected optically. The echo from the top layer returns to the surface first, followed in sequence by the echoes from each subsequent film interface. The time-evolved nature of the echoes provides time-resolved information about each layer, simplifying the analysis of multilayer film stacks.

As implemented in the MetaPULSE metal-film metrology system, for example, the technique is typically able to determine the thickness of each layer with angstrom-level accuracy and subangstrom repeatability for MLM film stacks composed of layers ranging in thickness from 2.5 nm to >3.0 µm. Because the nondestructive laser flash is focused on a small (~20-µm-diam) spot, the tool can utilize metrology test sites that already exist on product wafers. Depending on the thickness and complexity of the film stack, measurements take 1­5 seconds per site, which—assuming five measurement sites per wafer—provides a throughput of 40­60 wafers per hour. This capacity is sufficient for the technology to be used in in-line process monitoring.

Yield Management Applications

There are three reasons why the ability to perform MLM film stack analysis on product wafers can lead to improved yields. First, a single-layer metrology approach to characterizing MLM stacks typically requires that a monitor wafer be run for each deposition chamber (or layer) at least once per shift. For a cluster tool that deposits a three-layer stack at a rate of 30 wafers per hour, one set of monitor wafers per 10-hour shift would consume 1% of the output capacity. For product wafers having a value of $1000, this amounts to $3.25 million in lost production capacity per year for a fab operating at 5000 wafer starts per week. On-product MLM metrology has the potential to prevent this loss.

Secondly, the single-layer metrology approach requires an act of faith on the part of the process technician, who must make wafer disposition decisions based on a limited sampling of single-layer films processed at some time in the past. Although statistical methods are used to help minimize risk, out-of-control events seem to inevitably occur between monitor wafer runs. The best-designed statistical process control systems can (and frequently do) allow defects caused by process failures to go undetected until the affected wafers are completely processed. Not only do lots produced during these failures have to be scrapped, the time and money that was spent processing them further has been wasted. On-product MLM metrology improves process control engineers' ability to minimize the effects of process excursions and gives fab managers a new tool to help linearize fab output.

Finally, the process of diagnosing and correcting the causes of process excursions can be time-consuming and unpredictable. Conventional investigative tools such as transmission and scanning electron microscopy (TEM and SEM), while powerful, are slow and provide only a minute sampling of a single chip. In contrast, the measurement speed of the laser sonar technology can provide real-time multipoint analysis and uniformity mapping. In addition, few tools exist that can help engineers to assess factors not visible in TEM and SEM micrographs, such as film adhesion or material density. Based on extensive research, the developer of the laser sonar technology has compiled physical models and algorithms for calculating not only film thickness, but also such additional film properties as density, roughness, and adhesion strength from a wafer's echo-pattern data. In many situations, it is possible to obtain accurate values for these properties for individual layers even when measuring complex stacks. Many of the process excursions cited in the remainder of this article were recognized from PULSE measurements of film thickness plus an additional parameter. Most of the thickness measurements were subsequently confirmed by TEM, but in some cases pulse information alone proved critical to the rapid diagnosis of the processing problem.

Film Stack Integrity. In studies conducted in conjunction with the Sematech Analytical Lab Managers Working Group, a series of samples were generated and analyzed to determine the applicability of the laser sonar technology to MLM characterization. In one experiment, nominally identical six-layer MLM film stacks consisting of titanium nitride/titanium/aluminum-copper/titanium nitride/titanium/silicon dioxide (TiN/Ti/Al-Cu/TiN/Ti/SiO^₂) were deposited on two silicon wafers. The composition and order of these layers is typical of film stacks used in aluminum interconnect wiring. The Ti/TiN layers improve the adhesion between the aluminum metallization layer and the SiO^₂ interlayer dielectric insulator while at the same time providing low contact resistance to interlevel interconnect vias.

Examining the echo patterns reveals much about the film stacks. The first echo peak of each wafer occurred at approximately 10 picoseconds and the velocity of sound in most solids is in the range of 70 Å/ps. By inserting these values in the film thickness formula (sidebar equation 1), the thickness of the top film layer was calculated as ~350 Å. More-careful examination reveals that the echo in the first 50 picoseconds of Figure 2a is a decaying sine wave, which is typical of an echo reverberating in a single-layer film. The more-complex signal in the first 50 picoseconds of Figure 2b suggests that there may be more than one top layer in this film stack. For both samples there is a long quiet period from 40 to 130 picoseconds; this is what would be expected if sound were traveling through a very thick central layer. Using the thickness formula, the 120-picosecond period from the first echo at 10 picoseconds to the subsequent echo at 130 picoseconds indicates the presence of a central layer approximately 4200 Å thick. The echoes after 130 picoseconds are complex, suggesting the presence of multiple thin layers underneath this central layer.

Figure 2: Echo patterns from two wafers (a and b) on which nominally identical six-layer TiN/Ti/Al-Cu/ TiN/Ti/SiO2 film stacks were deposited. The red traces represent the echoes detected by the metrology tool and the blue traces represent the best fits of the multiparameter model used to calculate film thickness.

The thickness of each layer in the MLM stack on both samples was also determined using modeling. As the results in Table I reveal, the thicknesses of corresponding layers on the two wafers were comparable, with one exception: Wafer A did not have a Ti layer in its upper adhesion stack. This missing layer accounts for the somewhat early arrival of the echo from the bottom of the thick central layer on this wafer. These thickness values were confirmed by, and correlated well with, x-ray fluorescence (XRF) spectrometry, refractive back-scattering spectroscopy, and other measurements made on single-layer monitor wafers processed in the same multichamber deposition tool used to produce the sample MLM film stacks. It was then revealed that for this evaluation, the Ti layer in the top adhesion stack had been intentionally omitted from Sample A to challenge the laser sonar technology's ability to detect misprocessing in complex MLM film stacks. The technology clearly met this challenge.

Buried-Layer Uniformity. A second experiment demonstrated the ability of the laser sonar technology to detect accidental misprocessing by analyzing the uniformity of buried layers in an MLM film stack on product wafers. In this case, a fab's beginning-of-shift resistivity analysis of monitor wafers from a cluster tool that produced an Al-Cu/TiN/Ti/ SiO^₂ stack had indicated that the Ti-layer deposition had become very nonuniform at some time during the previous shift. As a result, all of the wafers produced after the previous (good) qualification wafer set had been pulled from production.

To troubleshoot this problem, representative product wafers from suspect lots were analyzed using the new metrology. Figure 3 shows a map of the buried Ti layer on one of these wafers. The map was created by taking measurements of the Al-Cu/TiN/ Ti/SiO^₂ stack at the same location on each of the 55 dies on the wafer and then plotting only the thickness of the bottom Ti layer. The results indicated that this thickness varied by more than 55%, from a minimum of 251 Å to a maximum of 440 Å. All other layers in the MLM stack were found to be uniform and to have the expected thicknesses.

Figure 3: Film-thickness uniformity map of the bottom Ti layer in an Al-Cu/TiN/Ti/SiO2 film stack, which was produced by taking one MLM measurement on each of 55 product dies. The arrow indicates the notch position.

These data pointed to a significant problem with the Ti-deposition chamber. Because it is unlikely that a gas-flow problem could account for this degree of nonuniformity, a physical problem was suspected and the chamber was disassembled. It was then discovered that the magnet that normally focuses the sputtering plasma onto the Ti target had fallen from its mounting bracket onto the target, where it partially blocked and distorted the beam. Remounting the magnet in its proper position restored the normal operation of the deposition chamber. In addition, individual screening of the remaining wafers in the suspect lots was performed with the laser sonar tool to identify those that met specifications, allowing many wafers to be saved.

Barrier and Seed Layer Variations. Semiconductor manufacturing is rapidly moving from aluminum to copper interconnect wiring in order to take advantage of the latter metal's lower resistance, which, in combination with low-k dielectric materials, can lead to higher processor speeds and lower power consumption. The well-documented mobility of copper and its ability to poison transistors, however, require that an effective barrier layer be included in copper interconnect MLM stacks. Tantalum nitride (TaN) films have been shown to provide such a barrier to copper diffusion while at the same time promoting good adhesion between the copper lines and the surrounding interlevel dielectric (ILD).

In a common processing approach, the TaN barrier layer and the first layer of copper (the seed layer that becomes the cathode for subsequent copper electrodeposition) are deposited in a single PVD cluster tool. As a result, it is important to have an MLM metrology system that is capable of determining the uniformity of both layers in the completed film stack. Such a tool can be used to optimize the seed/barrier-layer deposition process as well as to monitor production.

In one fab, the laser sonar technology was used to determine the thickness of the seed and barrier layers at each point in a 49-point wafer map as part of optimizing a new Cu/TaN process. The thickness of each layer also was determined at points along a radial line running from the center to the edge of the wafer. Between the wafer's center and a point 10 mm from the edge, measurements were made every 10 mm. In the outer 10 mm of the scan, thicknesses were determined every 500 µm along the line. The measured thicknesses of both films along this line scan are shown in Figure 4. The thickness data at one measurement site were confirmed by TEM, but because of the time and expense of the analysis, TEM was not performed at other wafer sites.

Figure 4: Measured thickness of a 300-Å TaN barrier film under a 1200-Å copper seed layer.

The line scan indicated that the barrier film was approximately 10% too thick in the center of the wafer while the copper seed layer was approximately 10% too thin in that area. This variation was confirmed in the film-thickness maps for the two layers. The map for the barrier layer, shown in Figure 5, indicated that the TaN film was not well centered on the wafer and that in areas adjacent to and opposite the alignment notch it was less than 200 Å thick. An inadequate barrier in the edge die in just one quadrant of an 8-in. wafer can result in a 5% drop in yield. The uniformity map for the copper seed layer on this wafer also indicated poor centering, a condition that can lead to a poorly centered layer of electrodeposited copper. By instituting process changes based on these MLM measurement data, the fab was able to optimize the deposition process so as to maximize yields.

Figure 5: Uniformity map showing the film thickness of a TaN barrier film under a copper seed layer. The arrow indicates the notch position.

Reactively Sputtered Film Density. The properties of reactively sputtered films such as titanium nitride depend strongly on such variables as pressure, temperature, and gas-flow rate. Under improper growth conditions, both the effectiveness of the layer as a barrier and its electrical conductivity may degrade significantly because of such factors as low density (excessive porosity) and low nitrogen content. In cooperation with process engineers from a fab that had experienced problems related to TiN-layer deposition, the ability of the laser sonar technology to detect variations in film density was studied.

The metrology tool was used to compare the echo patterns of two reactively sputtered TiN films, as shown in Figure 6. One film was prepared under conditions that produced a nearly stoichiometric composition, while the other was grown with a 50% higher nitrogen flow. Film density affects the rate of decay of an echo reverberating within the film, and it is clear from Figure 6a that the echoes excited in the stoichiometric film damped out at a substantially slower rate than those observed for the other film (Figure 6b).

Figure 6: Metrology tool measurement results for two reactively sputtered TiN films (a and b), which were produced under different conditions.

Calculating the film density from the damping rate using the theory of elasticity (sidebar equation 3) yielded a density for the stoichiometric film of 4.3 g/cm³, a value that falls within the expected range for a solid, continuous film. For the film using the higher nitrogen flow rate, however, the density was found to be 2.7 g/cm³, indicating that the film was highly porous. The thicknesses of the films (calculated from the time between echo peaks) were very similar at ~300 Å. Comparable results also have been obtained in studies evaluating the quality of reactively sputtered tantalum nitride and tungsten nitride films.

Mass-Flow Controller Failures. Malfunctioning gas-flow systems can adversely affect the thickness and other properties of reactively sputtered films embedded in MLM film stacks. When single-layer monitor wafer metrology is used for process control, such effects may go undetected for many hours, causing significant yield losses.

Another experiment was performed to help a fab diagnose such a problem and to demonstrate the ability of the laser sonar technology to detect mass-flow controller (MFC) problems. A series of MLM film stacks having the same nominal structure (TiN/Al/ TiN/Ti/ILD) were prepared, but during the growth of the bottom TiN layer, the nitrogen flow was varied from sample to sample to simulate normal operation and three possible modes of MFC failure: the controller opening late, the controller closing prematurely, and the controller failing to open. The resulting thickness data are summarized in Table II.

By comparing the thickness measurements for each sample with the data for the control, it is possible to determine the causes of the variations. When the nitrogen was turned on late, the Ti layer was very thick and the bottom TiN layer was thin. When the nitrogen was turned off early, the Ti layer was near the expected thickness and the bottom TiN layer was thin, but unlike the first sample, another thick Ti layer was deposited above this TiN layer. Finally, when no nitrogen was present, the TiN layer failed to form, and the Ti layer was very thick. If the laser sonar tool were used for the real-time monitoring of product wafers, achieving results similar to these would provide easily recognized indications of each failure mode.

Bonding-Pad Failures. After the interconnect wiring structure of an integrated circuit has been created, a passivation (or capping) layer composed of one or more dielectric films such as tetraethoxysilane (TEOS) is deposited across the top surface of the wafer to protect the chip from the long-term negative effects of ambient oxygen and moisture. In order to keep the IC connected electrically to the outside, the passivation film is masked with a patterned resist film that has small (~50-µm-diam) holes centered over each bonding pad. These holes allow windows to be etched through the passivation layer (as shown in Figure 7) so that wire leads can be welded to the MLM film stacks. Although this etch process is specifically designed to prevent etching of the MLM stacks, process problems can lead to difficulties in the subsequent wire-bonding process or to long-term failures of wires that are bonded to damaged pads.

Figure 7: Schematic representation of a bonding pad and the window that is etched through the passivation layer in order to allow a wire to be attached to the MLM film stack.

An unusual electrical signature obtained during testing caused a semiconductor manufacturer to believe that the MLM film stacks in the contact-hole windows in a large number of wafers had been unexpectedly etched. Confirming this overetch hypothesis was difficult because most of the analytical tools that were available in that fab were unsuitable for evaluating the condition of the MLM stacks within the 50-µm-diam window. Profilometry cannot indicate the actual thickness of an MLM stack, and other metal-film analysis techniques such as XRF are unable to take measurements within the small windows. Even if they could, signals from overlying layers and underlying levels would interfere, making it difficult or impossible to achieve accurate thickness measurements of metal 5-level bonding pads. The fab therefore decided to cross-section a small portion of a wafer through a bonding pad, perform TEM, and then analyze wafer samples with laser sonar technology.

The metrology tool's small measurement spot allowed the MLM stacks inside the windows to be easily analyzed without the need for time-consuming sample preparation. In addition, because of the time-evolved nature of the tool's signal, interference from underlying levels could be prevented by simply stopping the echo-recording feature before signals from these layers could return to the surface. As the blue trace in Figure 8 shows, the echo pattern from a bonding-pad window indicated that the top TiN layer was missing in that area. There was no echo in the first 10­20 picoseconds, and the first acoustical contribution to the signal did not come until about 150 picoseconds, which corresponds roughly to the round-trip time for sound for an aluminum-copper central layer having a thickness of approximately 4000 Å. Modeling analysis of the measured data gave thicknesses for the aluminum-copper and underlying TiN and Ti layers of 4128, 651, and 323 Å, respectively, as shown in Table III. Despite the fact that the top TiN layer was missing and the aluminum-copper layer was only 80% of the expected thickness value, the underlying TiN and Ti layers were within specification. This combination of results strongly suggested that the MLM film stack within the window had been overetched.

To confirm that overetching had taken place, an area of the MLM stack adjacent to the bonding pad that was covered by capping films was also analyzed. The echoes from this area, where no etching could have taken place, are shown in red in Figure 8. The early arrival time of the first echo (<10 picoseconds) and the period of its ringing indicated that the top TiN layer was present, and the film was 272 Å thick—close to the 250-Å specification. Sound waves from the underlying Al-Cu/TiN interface reached the surface of this top layer at 160 picoseconds, indicating that the aluminum-copper layer was 5163 Å thick, nearly 1000 Å thicker than in the center of the bonding-pad window. Analysis of the subsequent echoes indicated that the thickness of the TiN film was 714 Å and the thickness of the Ti film was 328 Å.

When, as in this instance, the laser sonar tool is used to measure an MLM film stack that is buried under transparent films, the thicknesses of the transparent layers can also be determined. Because the dielectric layers are transparent to the laser light, the sound wave is generated by absorption of the laser pulse in the top metal-film layer. This launches sound waves upward into the dielectric stack and gives rise to the oscillating baseline seen in the red trace in Figure 8. Modeling analysis of these oscillations gives the thicknesses of the Si^₃N^₄ and TEOS capping layers which, in this example, were found to be 3251 and 2997 Å, respectively, which are well within the specifications for these layers.

Figure 8: Echo patterns from two areas of the bonding pad shown in Figure 7. The MLM film stack in the etched window in the center of the bonding pad produced the echoes shown in blue, and the echo pattern produced by the film stack covered with TEOS and Si3N4 is shown in red.

Further investigation at the fab in which the wafers having the overetched bonding pads were produced revealed a faulty seal on a chemical etch tool, which had allowed water to get into the resist stripper solution. In the absence of water, the stripper solution does not etch metal. When water is introduced into the solution, however, a chemical reaction takes place that causes the contaminated solution to etch aluminum.

Conclusion

A new opaque-film metrology system enables process engineers to quickly detect yield-killing misprocessing events that affect the layers in MLM film stacks. Because the technique uses laser light and detects ultrasonic signals, real-time nondestructive measurements can be made on a wide variety of film stacks on in-process wafers at any stage of production. The thicknesses of transparent films that have been deposited on top of a metal film also can be obtained. In addition, the metrology provides information about several of the MLM films' physical properties, including density, roughness, and adhesion, which helps engineers to diagnose the root cause of various process line problems with a minimum of downtime and lost production. The technique also gives fab managers a powerful new tool to linearize production and minimize waste.

References

1. M May, "How the Computer Got into Your Pocket," Invention & Technology 15, no. 4 (2000): 46­54.
2. F Lakhani, DL Dance, and R Williams, "Design and Validation of 0.25 µm Integrated Circuit Yield Model," Semiconductor International 21, no. 6 (1998): 195­200.
3. H Maris, "Picosecond Ultrasonics," Scientific American 279, no. 1 (1998): 86­89.

George J. Collins, PhD, is the director of marketing at Rudolph Technologies (Flanders, NJ). He has more than 20 years of experience in several areas of analytical instrumention, including mass spectrometry, scanning probe microscopy, and thin-film metrology, publishing many papers in these fields. He received a PhD in food science and analytical chemistry from Rutgers University in New Brunswick, NJ, in 1980 and an MBA with a focus on marketing from Rutgers University in Newark, NJ, in 1990. (Collins can be reached at 973/448-4314 or gcollins@rudolphtech.com.)

MicroHome | Search | Current Issue | MicroArchives
Buyers Guide | Media Kit