Corrective Actions to Meet Extreme Tolerance Requirements for Thin Films: How to make peace with your deposition tools
By Robert Aigner, PhD, Director Acoustic Technologies R&D and Gernot Fattinger, PhD, Project Manager BAW R&D, TriQuint Semiconductor, Inc.

The article will examine the importance of tolerances as applied to the design and manufacture of ultra-precise RF components in an integrated circuit fabrication environment. Focusing on bulk acoustic wave (BAW) processes, the article discusses how the 3-sigma tolerance of thin film devices can be narrowed down to less than 0.05%. “Localized processing” (or “trimming”) using a scanned ion-beam approach allows correction of both run-to-run variations and non-uniformity in thin films thus eliminating the major yield inhibitor for bulk acoustic wave components. The article describes the principle and implementation of localized processing having a throughput suitable for high-volume manufacturing. Discussion will center around the challenges of adhering to precise tolerance specifications and the benefits derived from achieving these goals in terms of manufacturing yield and device performance. Practical examples will be shown. An outlook on potential future applications of localized processing in applications other than BAW devices will be presented.

Introduction
Absolute precision of components is often considered an expensive luxury in the electronics industry where tolerances of 5% to 10% are often thought to be adequate for general commercial applications. Premium prices are demanded for components having tight tolerances. Squadrons of engineers are engaged in finding solutions that are less sensitive to manufacturing tolerances. In IC manufacturing designers often get away with higher tolerances by relying on differential designs, by adaptive active circuit concepts or by entirely digital solutions. In RF systems the number of potential solutions shrinks drastically since passive elements cannot be avoided; they must often be low loss, highly linear and low noise. If those requirements exist the chances are low that a tunable element of any kind will work. The last resort is to introduce “trimming”, which involves a process to permanently alter physical dimensions of a component in a way that narrows the tolerances of a certain parameter in a desired fashion. The most widely used trimming method is laser trimming. It is applicable to resistors, inductors and certain capacitors. Each component is treated as an individual and processed sequentially; the process is therefore relatively slow and expensive. Moreover, there are applications where laser trimming is not suitable due to the nature of the devices to be trimmed.

RF filters based on thin film bulk acoustic wave (BAW) technology provide extreme selectivity, low losses and extremely narrow frequency tolerances. BAW filters outperform the classical surface acoustic wave (SAW) filters in applications where loss and selectivity are critical. In a BAW device, the acoustic wave propagates in the vertical direction just like in a quartz crystal. The main difference is that BAW filters operate typically above 1 GHz while quartz crystals function below 50 MHz. The frequency at which the resonance occurs is determined by the thickness and mechanical properties of all layers involved. In BAW devices all layers are deposited using thin film processes. The thickness range for the layers in a 2 GHz BAW device is anywhere from 20nm to 1.5µm (Figure 1). The Piezoelectric layer used in most BAW process is sputtered Aluminum Nitride (AlN). Electrodes are made from materials such as AlCu, W, Mo, Pt, or combinations thereof. In the case of solidly mounted resonator (SMR) type BAW, the acoustic energy is trapped in the resonator by a Bragg reflector that is constructed from stacked thin films having alternating acoustic properties. SiO2 and W are excellent candidates. The thickness of each layer in the Bragg reflector is adjusted in a way that 99.999% of the acoustic wave energy is reflected at the operating frequency. Other layers used in BAW devices include seed layers and passivation layers such as Si3N4, TiN, Ti and Au. The complete layer stack of a BAW device can range anywhere from 10 to 20 distinct layers.

One of the big challenges for volume manufacturing of BAW filters is the tight tolerance window for the frequency position of the final product. Some applications require a 3s of less than 1 MHz for a filter at a center frequency of 2 GHz. This converts in to 0.05% relative 3s error. Even the most advanced thin film deposition equipment used in semiconductor processing cannot guarantee better than 1% compounded thickness error, more typically this error will be 5%. Contributors to the compounded error are non-uniformity across wafer, run-to-run variation and chamber-to-chamber mismatch. Moreover, there are at least 5 layers which contribute significantly to the total error (Table 1) and those errors accumulate according to statistical error propagation. On top of all this misery that designers must consider, there are variations of material parameters such as density or Young’s modulus, (Footnote: Young’s modulus [N/mm2] describes how much relative deformation (or “strain”) a material exhibits under the influence of unidirectional mechanical stress [N/mm2]) which can cause additional frequency errors. As a consequence it is a virtually hopeless endeavor to manufacture BAW devices without a suitable method to correct the errors in thin film deposition. The frequency spread of a 2 GHz BAW wafer is in a range of 40 MHz (Figure 2). A 150mm wafer can yield a total of 13,000 BAW filters, but only 500 filters will fall in the specified tolerance window of ± 1 MHz. Accordingly the maximum yield would be 5%; this is in the ‘lucky’ case when no other yield losses occur.

Ion milling has a long tradition in trimming quartz crystals and enables achieving a tight frequency tolerance. In this process Argon is ionized and accelerated in an electric field to an energy of 400-1500 eV. The Ar+ ions hit the surface and knock out material from it, an effect that resembles sandblasting, but on the atomic level. The ion milling process requires components to be in a vacuum chamber. Pure physical etching with ion bombardment is applicable to all kinds of thin films, even materials which are hard to etch with dry chemistry (Cu) or chemically inert (Pt, Au, etc.) In principle such a process would work for singulated BAW devices, but it would take an excessive amount of time to handle and trim the 13,000 chips that fit individually on a wafer. This approach disqualifies ion milling from consideration as a tool for volume production work.

“Localized processing” as a Solution for Volume Production
Localized Ion Beam etching is a method used to correct surface errors in ultra precise optical lenses and mirrors such as those used in lithographic equipment and telescopes. A narrow ion beam with a Gaussian intensity profile is scanned along the surface with controlled speed and will remove the desired amount of material at each location. In order to utilize this concept for semiconductor processing the tool must be fully automated with a robotic handling system and a high-performance x-y scanning table carrying a cooled wafer chuck. The Ion Beam source must include a neutralizer to avoid charge accumulation at the wafer surface. The FWHM of the Gaussian beam is in a range of 10- 15 mm. The scan pattern is usually a meander with a line-to-line distance of 2-6 mm. The ion beam diameter is small enough to correct the thickness gradients which typically occur on a length scale of a few millimeters, but it is much larger then one BAW device. The trimming process works “region by region” rather then “device by device” (as in laser trimming), which is an important advantage for throughput.

In order to correct a thin film after it is deposited the thickness is mapped using a metrology tool suited to the type of film. Choices for metal films range from x-ray based tools to laser-acoustic tools. Transparent films are usually measured with optical spectrometers. The number and position of points in the map must be wisely chosen to reflect the true profile and avoid over- or under-representation of edge regions. Often a 49-point map will suffice. Sophisticated interpolation and sanitation or ‘scrubbing’ of the map data is important to remove errant data. The resulting thickness error map, the ion beam profile and the predefined scan path are input parameters for an iterative software algorithm that performs a de-convolution, which then yields a map of residence time for each coordinate along the scan path. Residence time (Figure 4) can be translated into scan velocity. De-convolution is the inverse operation to convolution (which is an integration of local dose at each location over time).

In a relatively uniform layer with a smooth profile exhibiting small local gradients the error map and the residence time map look virtually identical. The only limitation for such a case is the maximum velocity at which the scanning system can travel because it determines the minimum removal. The situation changes dramatically for error maps with strong gradients. The de-convolution roughens the residence time map severely. The accelerations required to accommodate the resulting velocity profile can be brutal, in fact a narrow beam is useless if the acceleration limits the ability to make steep gradients. For a given etch rate and beam size the maximum scan velocity and acceleration determines whether a certain thickness error can be corrected at all. The velocity limitation can be circumvented by lowering the etch rate or by adding more offset to the deposited thickness and then remove more material. In both cases the throughput of the system will plunge. The other thing to keep in mind is that the method relies on a stable etch rate and beam shape. Any error in etch rate or beam shape will compromise the results, even more so if more material has to be removed.

A Faraday cup is used to measure the local current density, which the incoming ions constitute and thus verify that the beam is stable. A full calibration of the beam profile requires test wafers to be etched with a certain test profiles and mapped with high resolution before and after the ion beam correction. As the effort for a full calibration is high it is important to extract as much SPC data as possible from the Ion beam source, correlate it with processing results, and create a feedback loop. An example of a test profile is shown in Figure 5. This test profile was processed into a blanket film of 100nm thick PE-CVD Si3N4 and the error in removal was evaluated.

Specific Requirements for BAW Devices
In BAW devices it is possible to correct for frequency errors caused by thickness variations of several layers in one single correction step, for example by ion Beam etching of the top electrode. The input data will be derived from on-wafer RF probing of the BAW devices. While RF probing is by far the most accurate way to determine an error map it is not trivial to calculate the amount of removal required to correct the error. This is because the sensitivities shown in Table 1 are interdependent in a nonlinear way. As a consequence it is important to make an educated ‘guess’ concerning how the individual layer profiles behave and take this into account.

As mentioned before ion beam etching is able to affect virtually every material. However, some prominent materials used in integrated circuit production (such as aluminum) are more difficult to deal with than others. Like most metals (with the exception of the noble metals) aluminum has a tendency to grow a native oxide layer when exposed to air, the thickness of which depends on exposure time, humidity and temperature. Unfortunately the ion beam etch rates of Al2O3 are a factor 5 times lower than those of pure aluminum. As a consequence when etching Al and AlCu alloy thin films the removal is initially slow but speeds up tremendously when the native oxide is gone. This problem makes it very difficult to obtain accurate removal. In contrast to aluminum, the thickness correction in tungsten and molybdenum works very well despite the existence of native oxide on those metals. This is because the removal rates of the native oxides matches those of the bulk metal very well. Dielectric layers do not exhibit this problem at all. It is therefore beneficial to have a thin Si3N4 layer on top of the final resonator just for the purpose of accurate trimming. The sensitivity of a Si3N4 layer at the top of the structure is small. On the positive side this enables very accurate trimming, but on the negative side the trimming range is very limited. It is highly beneficial to do a 1st “rough” trimming step in a layer with a high sensitivity and a 2nd “fine” trimming step in the topmost layer utilizing a low sensitivity. In a typical 2 GHz resonator the minimum removal will be in the low nm range, the maximum removal can be up to 30nm. The process time for one wafer ranges from less than five minutes to as much as 15 minutes. Processing time is strongly dependent on the type of profile and volume to be removed.

Ion beam etching is compatible with conventional photoresist on wafers used in acoustic wave device manufacture. This makes it possible to do lithography-defined trimming of certain groups of BAW resonators, which is very important for bandwidth adjustments in these devices.

Tools for ion beam trimming are commercially available from two vendors, both of which are well established with proven manufacturing abilities: (Epion/TEL www.epion.com and Roth&Rau www.roth-rau.de). The amount of data to be collected for each wafer is substantial and by far exceeds what semiconductor manufacturers normally collect for statistical process control (SPC). The infrastructure required for trimming includes wafer ID at each process step and upload/download of data. Each and every wafer is treated as an individual throughout the process. Once the infrastructure exists the trimming process adds insignificant complexity and cost to the manufacturing flow.

Dissemination of Localized Processing into Other Applications
There are a number of processes which can and will benefit from the ability to do localized processing. Obvious examples include MIM capacitors for which the thickness tolerance of the dielectric material can be significantly narrowed. Thin film resistors can be trimmed to tight tolerances. The magnetic layers in read/write heads for hard drives also benefit from thickness corrections that can improve performance and yield significantly.

Summary
Introducing the method discussed above allows the achievement of extremely tight tolerances for a variety of thin film processes without abandoning established deposition equipment and methods. The benefits for yield and performance are undeniable and will pay for the cost of processing in a very short time. Developing this process in a manufacturing environment could be one of those rare cases in which engineers who seek to exploit the benefits of creating devices with high tolerances and accountants who tend to favor reducing costs agree that the investment is justified by the return. It is possible to make both groups happy ….

TRIQUINT SEMICONDUCTOR
www.triquint.com
TXTLINX.COM105
Email this article to a friend!