Minimum Area Required for Poly Etch Endpoint Detection

Endpoint detection used in plasma etching is a very important process control technique.¹ It is a method that identifies when the film of interest has been removed. Detecting the endpoint, by design, ensures a consistent amount of etch on each wafer. Endpoint detection prevents under-etching caused by low etch rates or thick incoming films and over-etching caused by high etch rates or thin incoming films.

Tool-level statistical control can utilize the time-to-endpoint to monitor the etch tool stability. The endpoint time on each wafer can be compared to a reference distribution. If the endpoint time on a given wafer is outside the normal reference distribution, the plasma etch tool can be stopped to allow engineering intervention to prevent further material impairment. The shape of the endpoint trace can also be analyzed or compared to a reference endpoint trace to further monitor the quality of the process.² Analyzing the shape of the endpoint trace is useful in identifying incoming problems on the wafer, such as the wrong film stack, as well as problems with the etcher, such as a degradation in etch rate uniformity or a change in selectivity to the stopping layer.

There are numerous methods of endpoint detection. Single-wavelength optical emission is the most common. Other methods include multi-wavelength optical emission,³ plasma impedance change,⁴ DC bias change, the floating potential of a Langmuir probe,⁵ and diode laser absorption spectroscopy.⁶ This article uses the single-wavelength optical emission method.

There have also been many methods developed to analyze the signal being monitored to detect the endpoint. Most of these methods have been statistical,⁷ derivative or combinatorial⁸ techniques to pick out a minuscule signal change in a background with high noise (i.e., very low open area). This article uses the delay/normalize/ threshold endpoint detection method. The delay is the time period when the endpoint signal is ignored. This allows for plasma stabilization during the early stages of the etch. The normalization is the time period over which the endpoint signal is averaged to establish the reference signal level. The threshold is how much the signal must change from the normalized value before endpoint is declared.

There have been somewhat arbitrary design rules implemented over the years dealing with how much open (non-resist covered) area is required on a product layout at a given layer to ensure robust endpoint detection. Rules have been used at the typical layers that use endpoint detection, such as polysilicon (poly), isolation, metal or capacitor. The poly layer endpoint detection is discussed in this report. This subject is touched on from the point of view of making photomasks in reference,⁹ but no data on actual loading effects or minimum open area is given. Until now, there has been no published information on how to measure the minimum open area required to detect endpoint. This article presents one such method and gives the results on two different poly plasma etchers.

There are two elements to the minimum open area design rule. The first and simplest is how much total open area in the reticle field (including the scribeline) is required to ensure there is a sufficient light signal change when the film clears to reliably identify the endpoint. The second and more difficult element is how to distribute the open area to prevent microloading effects like profile changes, CD changes, stopping layer punch-through, etc. This report gives data to address the first element only; that is, the minimum open area required for reliable endpoint for polysilicon etching.

The wafer size was 150 mm with a major flat and no minor flat. Blank stepping fields were used to obtain the desired range in percent open area. The blank fields started at the center of the wafer and radiated outward. Adjacent stepping fields were overlapped, so the fringing effects associated with using the stepper blades to mask the exposure were on the outside perimeter of the cluster of fields only and not between every stepping field. Thus, the variable open area contribution as the resist eroded at the resist fringes was minimized. This technique can be applied to metal and oxide etching, as well as poly etching. Figure 1 is an example layout, and Table 1 shows the targeted percent open area.

1. An example of the layout used to obtain the desired range in percent open area.

POCl₃ -doped poly wafers on 400 Å thermal oxide were used for the tests. Arch Chemicals' 906 i-line resist was used for patterning, and an ASML 5500/200 stepper was used to expose the wafers. Each stepping field was ~100 mm².

Two etchers were tested using a typical poly main etch process on each etcher. A Lam 4400 and Lam TCP 9400SE were used. The Lam 4400 configuration was 150 mm, unclamped, top powered, movable gap with a 520 nm bandpass filter and a photodiode detector. The Lam TCP 9400SE configuration was 150 mm, electrostatic chuck, top power to TCP coil, bottom power to the chuck, two-piece gas injection ring, quartz focus ring, 520 nm bandpass filter and a photodiode detector. The recipe used in the Lam 4400 is shown in Table 2, and the recipe on the Lam 9400 is shown in Table 3 .

The etch chambers had recently been cleaned, and the quartz endpoint detector windows were cleaned at the same time. The detector gain was set to a point of reference by running a poly etch plasma on a bare silicon wafer and adjusting the gain potentiometer to obtain an output of 5.6-6.9 Vdc.

Each wafer was etched well past its endpoint in order to see the full signal change when the entire poly was etched from the wafer. While etching each wafer, the detector signal was collected through the SECSII communication port on the etcher by a PC with a commercially available software package from Brookside Software (San Carlos, Calif.). Through the software, the light signal vs. time can be displayed and analyzed offline.

Brookside endpoint signal traces were collected for each wafer. The Brookside raw data are displayed above. Figure 2 shows the Brookside data for the Lam 4400, and Figure 3 contains the Lam 9400 data. The Brookside plots are simply an arbitrary number on the Y axis proportional to the endpoint detector voltage, which in turn is proportional to the plasma light intensity at 520 nm. Plotted on the X axis is time. Ten wafers were run on each etcher, and should be discernible in the Brookside plots. The first and last wafers were blanket poly wafers. Those wafers represented 100% open area and were replicated to check the reproducibility of the experiment. The rest of the wafers were run in order of increasing percent open area, with the one exception noted on the Lam 9400 Brookside plot.

2. Brookside plots are an arbitrary number on the Y axis proportional to endpoint detector voltage.

3. The arbitrary number on the Y axis proportional to endpoint detector voltage is proportional to plasma light intensity at 520 nm.

The goal was to determine two light intensity signals: a high signal before endpoint, called the normalized signal (Sn), and a low signal just past the peak when the signal is dropping but not yet bottomed out, called the endpoint signal level (Se). The difference between those two signals divided by the normalized signal is the percent change in light intensity at endpoint (%Δ, Equation 1).

The %Δ subtracted from 100 (Equation 2) is the threshold (T), which is commonly used in endpoint detection algorithms. When T is reached, the etcher automatically "triggers" the endpoint and proceeds to the next step in the recipe.

%Δ = ([Sn-Se]/Sn)*100     (1)

T = 100-%Δ     (2)

The normalized light intensity was estimated by finding the maximum intensity just before endpoint, then backing up to note the intensity at 10 seconds before endpoint. The signal at the midpoint of the drop in light intensity was initially picked to serve as the "endpoint" signal level. The midpoint proved to be too difficult to get a true value of the light signal because the Brookside data is collected every second and the poly fully clears in ~10 seconds, so ±1 second near the midpoint makes a sizable difference in the signal level. For this reason, the signal minimum was chosen as the "endpoint" signal level. This must be taken into account in the final analysis and conclusions.

Figure 4 shows the percent signal change vs. percent open area for the 4400 and 9400 etchers. The 9400 has a linear dependence, and the 4400 is quadratic. The reason for the difference is not certain. The 9400 has a heated endpoint window, and the 4400 does not. The 9400 also uses a chamber plasma clean process between each wafer, and the 4400 does not. Both of these differences would suggest that the chamber walls and endpoint detector window remain cleaner in the 9400 than in the 4400. It is also likely that the greater the open area on the wafer, the greater the deposit of etch byproducts on the chamber walls and detector window in the 4400. Regardless of the reason for the difference in the shapes of the curves, the region of interest in both cases is at the linear portion of the curves near the origin.

4. The Lam 9400 shows a linear dependence to the percent signal change vs. percent open area, while the 4400 is quadratic.

Another set of data derived from this test was the etch rate as a function of open area. Figure 5 indicates that the etch rate was independent of the percent open area for both the 4400 and 9400, which is surprising. In fact, there appeared to be no difference in etch rates between resist-coated wafers and uncoated wafers. This goes against conventional wisdom that poly etch has some chemical component, and an etch rate loading effect is expected. The absence of a loading effect could be explained in two ways:

1. There is very little chemical component in the etch mechanism, and the rate is dominated by ion-assisted etching characteristics (ion density and ion energy).
2. It does have a significant chemical component, but there is sufficient reactant available independent of the open area, and the endpoint is limited by surface effects (adsorption, diffusion or desorption) or the reaction kinetics between the poly and primary etchant.

5. A surprising find, the etch rate was independent of the percent open area for both etchers.

Using a simple threshold method of endpoint detection, the minimum drop in light signal that can be used to detect endpoint is 5%. From the curves in Figure 4 , a 5% drop on the 4400 occurs at 10% open area, and on the 9400 at 13%.

Recall that the 5% drop was taken from the minimum signal (i.e., after the entire poly cleared across the wafer). In a poly etch, it is desirable to call the endpoint before the poly fully clears, so the recipe can switch to a step with greater selectivity to oxide. A 10% drop to the minimum signal would roughly be the same as a 5% drop to the midpoint. On the 4400, a 10% drop in signal occurs at 22% open area, and on the 9400, it occurs at 27% open area. Tool-to-tool variations should also be taken into account by adding another few percentage points. Furthermore, there is always a small percentage of the wafer area that is unpatterned because of the field stepping pattern on the wafer, as well as global alignment marks (if used). The total unpatterned area is only 1-2%, which is spread over ~125 stepping fields per wafer — therefore inconsequential, and no correction for that effect is needed.

The final outcome is that the minimum open area for poly etch endpoint detection using the threshold method on the Lam 4400 is ~25%, and on the Lam 9400 30%. The design rules can be constructed to reflect the outcome of this test. If there is not enough open area, one of the more sensitive endpoint detection methods previously mentioned could be investigated.