The work discussed in this article treated the space bounded between the wafer and the pad as a reaction chamber having its own slurry input and output streams. In order to gather critical information about slurry concentration near the wafer surface, experiments were performed to determine slurry mean residence time (MRT)—the average time it takes for fresh, incoming slurry to displace an existing fluid in the reactor. Gathering that information is vital for several reasons: First, understanding slurry concentration and its dependence on polishing time is critical in process kinetics modeling. Second, transient analysis can be used to determine the effectiveness of water in removing slurry during subsequent rinse steps. Third, understanding the time constants associated with fluid replacement during CMP can enable the controlled introduction of special additives to quench or enhance polishing during multistep polishing on the same platen. Information on MRT will become even more critical as polishing times continue to decrease in accordance with long-range roadmap specifications. The overall goal of this study was to investigate the fluid dynamics and inherent tribological aspects of the CMP process in order to develop robust planarization processes with significantly lower slurry consumption.

To determine how much interlayer dielectric (ILD) is removed during a 30-second polish process using a silica slurry with a solids content of 20% by weight, as presented in curves (a) and (b) in Figure 1, it is typically assumed that the wafer is subjected to 20% slurry for the entire polishing time (t_p), as presented in curve (c). That assumption would hold true if the slurry replaced the water in the reactor instantaneously. In reality, transient conditions that depend on the relative magnitude of the polishing time and the slurry MRT must be accounted for in any removal rate model, as demonstrated by curves (d), (e), and (f). For example, curve (f) shows that the slurry concentration never reaches 20% during the 30-second polishing period, and that because of the long MRT, the slurry tends to stay in the reactor longer than 30 seconds.

Figure 1: Various scenarios of slurry output transients encountered during a 30-second polishing process using slurry with a solids content of 20% by weight: (a) water input to system; (b) slurry input to system; (c) MRT = 0 (ideal), (d) MRT < tp, (e) MRT = tp, (f) MRT > tp.

Such long transients influence the amount of ILD removed in at least two ways. First, because ILD removal rate depends on the amount of solids in the slurry, the slurry on the wafer surface undergoes extended dilution during the transient period. Second, the presence of extended transients allows the slurry to stay between the wafer and the pad long after the slurry source has been shut off, causing the polishing process to continue during the rinsing step. Such complex dependencies necessitate a fundamental understanding of MRT and the factors that influence it during polishing.

Based on these considerations, the work described in this article was divided into two broad phases. The first part of the study (Phase I) used a new method for determining MRT and explored the effects of key processing parameters on MRT based on classical chemical engineering reactor design theories and tribology. The data obtained from that investigation were then used to determine how much slurry is actually consumed during a typical polishing process (i.e., slurry utilization efficiency). The second part of the study (Phase II) explored the effects of MRT on ILD removal and investigated how this information can be used to design new CMP processes that consume significantly less slurry than current processes without adversely affecting ILD removal rates.

To conduct the study, a 1:2-scaled version of a 472 polisher from Speedfam-IPEC (Chandler, AZ) was constructed. Table I shows the appropriate scaling factors for each parameter and a numerical comparison between the typical values of the scaled polisher and those of a full-scale 472 polisher. Assuming that the slurry's kinematic viscosity and the pad-wafer space were the same on both polishers, the researchers used a Reynolds number to scale the platen and wafer speeds (i.e., the relative pad-wafer velocity in the scaled model was matched to that of the full-scale model). The scaled polisher's platen-to-wafer diameter ratio and slurry flow rate normalized by the platen area corresponded to the values for the full-scale polisher.

The scaled polisher and its associated attachments are shown in Figure 2. Instead of a standard polishing head, a modified industrial-rated drill press that rotated and applied downward pressure was used as the wafer carrier. A carriage equipped with dead weights and mounted on a traverse provided variable pressure onto a gimbaled wafer carrier. The pad conditioner consisted of a 76-mm diamond-grit wafer spring-loaded onto the pad, exerting downward pressure ranging from 0.5 to 1.5 psi. The diamond-grit wafer rotated and swept independently across the pad radius with the aid of two stepper motors.

Figure 2: Scaled polisher with friction table and pad conditioner (left, side view) and with drill press and traverse (right, front view).

The polisher was placed on top of a friction table consisting of two parallel plates. The plates were allowed to move relative to each other in only one direction. A strain gauge was mounted between the two plates to measure the lateral force exerted by the top plate onto the bottom plate during polishing. Measurement was achieved by calibrating between voltage output and force.

All polishing parameters were computer controlled and monitored. In addition, the computer synchronized the friction table to the polishing process so that real-time friction data—crucial for determining MRT—could be obtained during polishing. For any given run, the coefficient of friction (COF) was determined by dividing the shear force by the normal force applied to the wafer. Shear force was determined experimentally from the calibrated voltage output of the strain gauge, and normal force was determined by multiplying wafer pressure by the wafer surface area.

Mean Residence Time Analysis. A new technique was developed to measure slurry MRT in the wafer-pad region as a function of relative pad-wafer velocity, slurry flow rate, and wafer pressure. The technique uses the residence time distribution (RTD) method, which relies on the change in shape of the transient response (known as the F-curve) to an instantaneous disturbance within the system (such as the sudden replacement of water flow with slurry flow).² The value of MRT is determined by mathematically manipulating the F-curve with the equation:

Before the researchers obtained the F-curve, a perforated FX-9 polyurethane pad from Freudenberg Nonwovens (Lowell, MA) was subjected to a 30-minute ex situ conditioning process using PL-4217 slurry from Fujimi America (Wilsonville, OR) with a solids content of 25% by weight. Pad conditioning was performed with a 100-grit diamond disk at a disk pressure of 0.5 psi, a rotational velocity of 30 rpm, and a disk oscillation frequency of 20 oscillations/min. Pad conditioning was followed by a 5-minute break-in with a dummy wafer. The rotational speed of the wafer was matched to the rotational speed of the pad, and the slurry was injected at the pad's center. In all cases, the pad was conditioned in situ during MRT data acquisition. Table II summarizes the key process conditions.

In the ILD CMP process using fumed silica slurries and the scaled polisher described here, increasing the abrasive concentration of the slurry from 2.5 to 25% by weight caused the COF between the wafer and the pad to decrease from 0.35 to 0.14.³ This phenomenon resulted from the saturation of available abrasive adsorption sites on the surface of the pad, causing an excessive buildup of abrasives and a subsequent increase in lubricity between the wafer and the pad.^3,4

The RTD method took advantage of the effect of abrasive concentration on COF to construct an F-curve. First, the system was allowed to reach steady state using PL-4217 with a solids content of 25% by weight. The system then switched instantaneously to a PL-4217 slurry with a solids content of 2.5% by weight, causing the old slurry to be replaced and allowing the system to reach a new steady state. Figure 3 shows a data output that is similar to curves (a) and (b) in Figure 1. Using that information, the F-curve could be extracted, and by applying the above equation, the MRT could be determined experimentally.

Figure 3: Actual response data showing COF versus time for slurry with a solids content of 2.5% by weight displacing slurry with a solids content of 25% by weight.

Figures 4 and 5 present MRT and COF companion plots. The range and magnitude values of the COF transient represent the first and second steady-state conditions reached by COF before and after the rise shown in Figure 3. (The bottom and top of each bar in Figures 4c, 4d, 5c, and 5d represent the first and second steady-state values, respectively.) Based on approximately 20 repeat experiments, the relative standard deviations for MRT were estimated to be 10% while the relative standard deviations for COF were estimated to be 5%.

Test results presented in Figures 4 and 5 indicate that MRT ranged from 9 to 34 seconds, depending on the parameter settings. Both figures show that MRT decreased when slurry flow rate increased, regardless of wafer pressure and relative pad-wafer velocity. According to basic chemical reactor engineering principles, doubling the flow rate should result in a twofold reduction in MRT in a flow system of constant volume. However, in the case of a CMP reactor, perfect inverse proportionality is not observed for two reasons. First, the amount of slurry introduced onto the pad is not linearly proportional to the amount of slurry actually entering the wafer-pad region, since some amount of slurry falls off the pad, depending on the location of the injector, the design of the retaining ring, and the size and rotational velocity of the pad and wafer. Second, high slurry flow rates tend to cause the wafer to experience partial hydrodynamic lift, violating the constant-volume assumption.^5,6

At low relative pad-wafer velocities, increasing wafer pressure from 2 to 4 psi resulted in a 25% across-the-board increase in MRT, as shown in Figure 4. This increase was thought to be the result of compression of the pad under the wafer at high applied pressures. That compression increased the contact area between the wafer and the pad and increased resistance to slurry flow in the pad-wafer region. The data in Figure 4 indicate that COF values also experienced a 25% increase when pressure increased twofold. The increase in both MRT and COF values at higher wafer pressure suggests that there was intimate contact between the wafer and the pad, as illustrated in the insets in Figure 4.

At high relative pad-wafer velocities, as shown in Figure 5, increasing wafer pressure results in only a slight increase in MRT. That interaction does not violate the theory that intimate contact between the wafer and the pad results in higher MRT values; pad compression resulting from increased wafer pressure has been shown to be minimal at high velocities, because the wafer experiences hydrodynamic lift.⁶ The COF results presented in Figure 5 corroborate that hypothesis.

Since the experiments were conducted in an orthogonal manner, cross-comparison of the data contained in Figures 4 and 5 also could shed light on the effect of flow rate and relative pad-wafer velocity at constant wafer pressures. At 2 psi, doubling the pad-wafer velocity reduced MRT by more than a factor of two, indicating that the combination of low pressure and high velocity resulted in the least resistance to slurry flow.⁶ While doubling the wafer pressure to 4 psi also led to a reduction in MRT, the drop in MRT (as velocity was increased) was less because of greater asperity contact between the wafer and the pad.

Slurry Utilization Efficiency Analysis. The experimental determination of MRT using the RTD technique can shed light on the amount of slurry actually participating in the polishing process. General reactor design theory states that MRT is related to the volume (V) and volumetric flow rate (q) of the reactor as expressed in the equation:

Therefore, for any experimentally determined MRT, in order to calculate the value of q (the actual slurry flow rate under the wafer at any given time), information about the volume of slurry entrained between the wafer and the pad must be known. The volume bounded between the pad and the wafer is directly proportional to the apparent distance between the wafer and the pad. Previous studies using the scaled polisher and process conditions described in this investigation showed that distance to range from 18 µm (at high pressures and low relative velocities) to 25 µm (at low pressures and high relative velocities).⁶ The parameters investigated in those studies were all within the range of parameters selected for this investigation.

For a 100-mm wafer under the conditions of Figure 4, an apparent wafer-pad distance of 18 µm corresponds to a bounded volume of 0.15 cm³. MRT ranges from 21 seconds at 80 cm³/min to 33 seconds at 40 cm³/min. Based on the above equation, the slurry flow rate under the wafer therefore ranges from 0.0072 to 0.0046 cm³/sec, respectively. Given the amount of slurry injected onto the wafer, the slurry utilization efficiency is calculated to range between 0.35 and 1.1%. Under the conditions shown in Figure 5, the slurry utilization efficiency ranges between 1.7 and 2.6%.

These startling results attest to the wasteful nature of the CMP process, demonstrating that conventional methods such as reducing slurry flow rate, while effective at reducing overall consumption, do little to increase the utilization efficiency of the process. Innovations in polisher design, pad grooving, and slurry reuse methods are essential to significantly improve slurry utilization efficiency.

A New Dimensionless Parameter and Its Impact on ILD Removal. Phase I of this study showed that the choice of slurry flow rate, wafer pressure, and relative velocity result in a wide range of mean residence times. Based on those findings, a new dimensionless parameter, which had an impact on process optimization decisions and ILD removal rates, was introduced. The parameter, called the turnover ratio (TR), represents the ratio of MRT to polishing time (i.e., the amount of time during which slurry is injected onto the pad). TR is expressed in the equation:

For this series of experiments, polish time was held constant at 30 seconds, allowing TR to vary with MRT by selecting different combinations of pressure and relative velocity. Figure 6 shows slurry concentration plots versus time for three different TR values. Interestingly, the curve corresponding to the lowest TR reached a maximum solids concentration of 20% more rapidly than its counterparts. Moreover, the curve corresponding to the highest TR never reached the maximum solids concentration during the 30-second polishing period.

Figure 6: Concentration curves (F-curves) corresponding to 30-second polishes at three different turnover ratio values. P (Pa) x v (m/sec) = 17,099.

To demonstrate the effect of TR on ILD removal, a method was devised in which MRT could be studied as a function of ILD removal rate independent of the combinations of pressures and relative velocities required to arrive at different MRT values. This method was realized by using a slurry that exhibits Prestonian behavior over a wide range of solids concentrations.⁷ (Preston's law states that removal rate is linearly proportional to wafer pressure and velocity.) By using such a slurry, different combinations of pressure (P) and velocity (v) could be chosen to alter MRT, as long as the product of P x v was kept unchanged.

To determine MRT and, therefore, the TR corresponding to the concentration curves in Figure 6, IC-1000 k-groove pads from Rodel (Phoenix) were used to perform the experiments. Each pad was conditioned and broken in using procedures identical to those described in Phase I of the study. Table III summarizes the key process conditions.

The researchers used Syton-OXK colloidal silica slurry from DuPont Air Products Nanomaterials (Carlsbad, CA). With a solids content of 20% by weight, this slurry was selected among various colloidal slurries after exhaustive tests showed that it behaved most like a Prestonian slurry over a wide range of solids concentrations. Unlike the solids content of the fumed silica slurry used in Phase I, the solids content of the colloidal silica slurry has a nonmonotonic relationship to COF, as shown in Figure 7.

Figure 7: Effect of the solids content of colloidal silica slurry on COF.

To determine the colloidal silica slurry's MRT, the COF was initially brought to steady state using ultrapure water. The fluid was then switched to colloidal silica slurry with a solids content of 20% by weight for a prescribed period of time, after which water was reintroduced into the system. Figure 8, showing the COF response to the fluid inputs, highlights the nonlinear relationship between COF and solids content. The data contained in Figures 7 and 8 were then combined to construct the F-curve in Figure 9, which represents slurry concentration versus time and is similar to the curves in Figure 6. Based on the data in Figure 9, MRT was calculated using the first equation presented above.

Figure 8: Actual response data showing COF versus time for a colloidal silica slurry, with a solids content of 20% by weight, displacing water during a 30-second polish.

Figure 9: F-curve, constructed from the data contained in Figures 7 and 8, showing solids concentration versus time.

Next, removal analysis was performed for a series of polishing conditions. For each test, initial and final oxide thicknesses were recorded to determine the actual amount of oxide removed during each test. In addition, the COF response was recorded in real time (1000 times per second for 30 seconds) and converted to a concentration response, as presented in Figures 7 and 8. The concentration response (i.e., the F-curve) was used to determine MRT as well as the expected amount of oxide removed. A three-step procedure was used to determine MRT and oxide removal.

First, removal rate data were obtained experimentally for a wide range of wafer pressures and relative pad-wafer velocities. The polishes were carried out at four different slurry dilutions of 2.5, 6.25, 12.5, and 20% by weight. In order to avoid concentration transients during polishing, the slurry was introduced onto the pad 15 minutes before the wafer and the pad were brought into contact with each other. Based on a modified Preston's equation for oxide polish,⁷ removal rate (RR) was modeled in terms of wafer pressure (P) and relative pad-wafer velocity (v) by using chemical and mechanical rate constants (k_c and k_m):

Figure 12 shows the effect of TR on the actual and expected amount of oxide removed for a 30-second polish. (The error bars represent one standard deviation.) The point corresponding to a TR value of zero was an ideal case obtained by allowing the system to reach steady state with a slurry solids content of 20% by weight before contact was made between the wafer and the pad. Results of Figure 12 are:

Figure 12: Effect of turnover ratio on oxide removal for a 30-second polishing process using colloidal silica slurry with a solids content of 20% by weight.

A Simple Case Study. To test their findings, the researchers performed a baseline ILD polishing process using the colloidal sillica slurry with a solids content of 20% by weight. The slurry flow rate was set to 60 cm³/min, the wafer pressure to 2 psi, and the relative pad-wafer velocity to 1.24 m/sec. Experiments indicated that the TR for this process was 0.30, corresponding to 495 Å of oxide removal over a 30-second period.

Alternatively, a process can be designed that increases wafer pressure to 4 psi, reduces relative velocity to 0.61 m/sec, and reduces slurry flow rate to 30 cm³/min. Based on the general trends shown in Figures 4 and 5 and the data presented in Figure 12, the alternate process would be expected to yield a TR value between 0.5 and 0.6 and an oxide removal level of approximately 500 Å over a 30-second period. Assuming that lowering the flow rate would not adversely affect the defect density quality of the polished film, the alternate process would represent a twofold reduction in slurry flow rate with no adverse effect on oxide removal.

To develop robust planarization processes with significantly lower slurry consumption than is currently the norm, a study was undertaken to characterize the fluid dynamics and tribology of interlayer dielectric CMP. The resulting data were used to quantify the extent of slurry waste and to determine the dependence of ILD removal rate on slurry transport characteristics.

The average time it took for fresh incoming fluid slurry to displace existing slurry in the pad-wafer region (defined as the mean residence time) was determined experimentally using a residence time distribution technique in conjunction with real-time frictional analysis. Using the classical definition of MRT in conjunction with data on the apparent distance between the wafer and pad during polishing, the amount of slurry actually participating in the polishing process was determined, from which slurry utilization efficiency could be calculated. Results showed the slurry utilization efficiency to be less than 3%, underscoring the wasteful nature of the CMP process.

Based on these findings, the turnover ratio parameter was introduced. An increase in TR resulted in a significant decrease in ILD removal rate because of long transients at dilute slurry conditions and because short polishing times prevent slurry from reaching its maximum concentration during the polishing process. By considering MRT trends and scaling to the value of TR, it was demonstrated that slurry flow rate can be reduced by a factor of two (compared with a baseline process) without adversely affecting oxide removal, resulting in potentially significant environmental and cost benefits.

The authors wish to thank Fujimi America and DuPont Air Products Nanomaterials for donating the slurries used in the tests discussed in this article and Freudenberg Nonwovens and Rodel for donating pads. The authors would also like to acknowledge the technical assistance provided by Lorenzo Lujan and members of the Tufts University Fluid Laboratory. They are also grateful for the financial support provided by the NSF/SRC Engineering Research Center for Environmentally Benign Semiconductor Manufacturing (centered in the Department of Chemical and Environmental Engineering, University of Arizona, Tucson).

References

1. A Philipossian, F Sanaulla, and K Lopez, "CMP Consumables Pricing and the First and Second Laws of Thermodynamics," in Proceedings of the Third Annual Workshop on Chemical Mechanical Polishing (Potsdam, NY: Clarkson University, 1998), 1–17.

Ara Philipossian, PhD, is the Koshiyama associate professor of planarization in the department of chemical and environmental engineering, University of Arizona (Tucson). His research is in the areas of planarization and postplanarization cleaning. Philipossian teaches courses on fluid flow, heat transfer, and IC processing. From 1992 to 2000, he served as the materials technology manager at Intel, where he was responsible for the development, characterization, implementation, and sustaining of new and existing CMP and post-CMP cleaning consumables, low-k dielectrics, and electroplating chemicals. He has authored more than 30 refereed journal publications and 75 articles in conference proceedings. He holds 12 patents in the area of semiconductor processing and device fabrication. Philipossian serves as the advisory board member of the Planarization and CMP Technical committee of the Japan Society of Precision Engineers and is an active member of the Electrochemical Society and the Materials Research Society. He received BS, MS, and PhD degrees in chemical engineering from Tufts University in Medford, MA. (Philipossian can be reached at 520/621-6101 or ara@engr.Arizona.edu.)