CMP Characterization and Process Development for Undensified PSG

by Roger Su

Assistant Development Engineer, Microlab, UC Berkeley

Introduction

In recent years, chemical mechanical polishing (CMP) has become an increasingly important step in device fabrication. Whether it is the damascene process or planarization in MEMS fabrication, the CMP process plays a crucial role in these applications. The recent purchase of a Strausbaugh, model 6EC, single wafer CMP machine by the microfabrication laboratory here at UC Berkeley allows the lab to offer to its members a CMP process comparable to industry’s. Since it is the goal of the laboratory to provide a reliable and repeatable CMP process, a fractional factorial experiment was conducted to characterize the CMP machine. This report summarizes and presents the major results of the experiment.

Experiment Setup

There are many variables in the CMP tool that can be changed to affect a polishing process. Among them we selected five variables, down force (DF), back pressure (BP), table speed (TS), spindle speed (SS), and slurry flow (SF), as the main parameters in our experiment design. The following figure illustrates how a CMP tool works and the five major parameters involved in the polishing process. These five parameters are the five main effects of the experiment. Two responses were calculated from the experiment, removal rate (RR) and percent non-uniformity (NU).

Figure 1 - Main Effects in the CMP Process

The experiment design is a 2 _IV ^5-1 fractional factorial experiment matrix. Two runs of this experiment matrix were run within a few days. Each run consisted of the fractional factorial runs plus two midpoint runs. The two levels and midpoint values for the five variables of the experiment are shown in Table 1 below.

Variable	Units	+	o	-
Down Force	psi	15	11	6
Back Pressure	psi	4	2.5	1
Table Speed	rpm	100	80	60
Spindle Speed	rpm	40	25	10
Slurry Flow	ml/min	150	100	50

Table 1 - Variables, Midpoints, and Levels of the Experiment

The actual run order of wafers was randomized and is listed in Table 2. The same order was kept for both runs and a dummy wafer was polished before each run.

Wafer Order	Run #	Down Force	Back Pressure	Table Speed	Spindle Speed	Slurry Flow
1	12	+	+	-	+	-
2	15	-	+	+	+	-
3	4	+	+	-	-	+
4	10	+	-	-	+	+
5	5	-	-	+	-	-
6	2	+	-	-	-	-
7	6	+	-	+	-	+
8	17	o	o	o	o	o
9	9	-	-	-	+	-
10	16	+	+	+	+	+
11	3	-	+	+	-	-
12	13	-	-	-	+	+
13	8	+	+	+	-	-
14	11	-	+	+	+	+
15	1	-	-	-	-	+
16	7	-	+	+	-	+
17	18	o	o	o	o	o
18	14	+	-	+	+	-

Table 2 - Run Order and Variable Levels for the Experiment

All other factors were kept constant in the experiment, such as: slurry type (Cabot’s D7000 oxide slurry), pad type (Rodel’s IC1000/SUBA IV composite pad), table temperature (30^oC), pad conditioning (in-situ).

The process wafers were prime grade, p-type, <100>, bare silicon wafers originally. They were deposited with undensified PSG in tylan20 using the standard PSG recipe. The initial PSG thickness was measured for all wafers at five sites (top, center, flat, left, right) using the nanospec. Then, the wafers were polished by the CMP tool. Table 3 shows a generic recipe:

Step	1	2	3	4	5
Time (sec)	15	5	5	60	15
Down Force (psi)	0	2	VAR	VAR	2
Table Speed (rpm)	VAR	VAR	VAR	VAR	VAR
Spindle Speed (rpm)	VAR	VAR	VAR	VAR	VAR
Back Pressure (psi)	-2	-2	-2	VAR	-2
Slurry Flow (ml/min)	VAR	VAR	VAR	VAR	0

Table 3 - Generic CMP Recipe

used by the Strausbaugh CMP machine. Each wafer used this same recipe except that the fields with “VAR” have the values from the experiment design matrix for that wafer. The recipe has five steps. The first step starts the rotation of the table and spindle and spreads the slurry onto the table. The second step brings down the polishing arm at a low down force of 2 psi. The third step then increases the down force to the desired value. The fourth step is the main polish step where the back pressure is set to the desired value. The fifth and last step is a buffing step where water is used to give a final planarization to the wafer. After polishing, the wafers were cleaned and measured again by the nanospec at approximately the same sites for their final PSG thickness. The initial and final thicknesses for each wafer were recorded.

Results and Discussion

From the experiment runs, two numbers were calculated for each polished wafer: undensified PSG removal rate (RR) and within wafer % non-uniformity (NU). They were calculated from the following formulas:

For removal rate, the formula just calculates the average difference between initial and final PSG thickness from the five sites on the wafer. The unit for removal rate is angstroms removed per minute. For within wafer non-uniformity, the formula calculates the non-uniformity of the film after polishing. The reason for using this formula for non-uniformity requires a little more thinking. At first, we used the non-uniformity formula for plasma etchers which is essentially a measure of the variation of etch rates on the wafer. This is acceptable for etchers because topography is expected after etching and thus uniform etch rates are desired. However, the expectation for a CMP machine is just the opposite. A major purpose of CMP is to planarize the wafer surface. Ideally, there would be no topography on a wafer after CMP. Thus, the uniformity of a CMP process is indicated by the uniformity of the film after polishing.

The removal rate and non-uniformity data for the two experiment runs are presented in Table 4. The two repeated runs gave some repeatability to the result and ensured enough degrees of freedom for statistical analysis. The data from the two runs were combined and analyzed using JMP, a software program designed for DOE analysis. Since this was a 2 _IV ^5-1 experiment design, all main factor and two-factor interaction effects were resolved. The JMP program computed these effects and fitted a linear model equation to predict either the removal rate or non-uniformity. Each model equation was then refined to contain only the effects that were, with about 90% confidence level, statistically significant. JMP also calculated a R² value for each model equation which indicated how accurate the data fit the model.

	Run 1		Run 2
Run #	RR (Ang/min)	NU (%)	RR (Ang/min)	NU (%)
1	8546.2	10.38	7798.2	3.51
2