MEMS integration carries advantages, risks

MEMS integration carries advantages, risks
Is integration a viable approach for microelectromechanical systems (MEMS)? Should a MEMS project select the path of single-chip integration even if it requires additional time and risk? Perhaps putting several dice in the same package is a better solution, but what criteria should be used to decide? The answers are not easy to determine: The variables are difficult to quantify, and often success can be determined only after a product is in the market.

The timing risks associated with extended learning cycles and complex device and process interactions have made integration uncommon in micromachining. This paper will explore MEMS integration and also look ahead to future trends when functionality and volumes could change some of the assumptions commonly applied today for MEMS-based products.

When IC chips were a few transistors (as in the early days of MOS technology), the technology had what seemed to be unidentified variables that were more like random numbers than well-behaved functions. The tools to define early planar devices were simple graphics and models of the process. The device characteristics were commonly created with math approximations and slide rules. The 100 million-transistor IC was proven more than once to be impossible based on nothing more than yield and defect density.

In hindsight (as we use our 100 million-transistor Itanium processor chips), the improvements that have taken us along Moore's Law were progressive steps in the understanding of design and manufacturing tools-tools that minimized the risk of failure. Single-chip integration of ICs seems to have won almost all of the arguments over the years, with the well-deserved conditions that dictate the best engineering solutions. With its high performance at a low price point, and high levels of reliability and quality, integration has been the key business differentiator of all the IC companies that have survived.

As with ICs, the first situation to consider with MEMS is whether the volumes are large enough to integrate and thus amortize the development and tooling costs. Then the question of a viable process must be answered. The most compelling argument for integration is whether there's no other way to get to the low price targets in volume. The decision can often be finalized when the physical-size limits dictate that nothing else is feasible. The quality and reliability levels that are backed up by warranty and recall implications, multiplied by the large volume, might influence any trade-offs. After fielding more than 120 million single-chip MEMS devices, Analog Devices' experience confirms that defect levels below 1 ppm are routinely possible and mean time to failure has been greater than 1 billion hours as a result of single-chip MEMS integration.

In terms of technical solutions, putting all the elements on a single chip has some distinct advantages. Small signals in the presence of high levels of noise are best handled with a minimum of unknown variables derived from stress, electromagnetic interference, parasitic capacitance and electrical leakage. The common techniques possible with IC design, such as cross-quading and switch-capacitor charge management, make it possible to cancel effects due to temperature and other unknowns.

Integration puts the electronics together with the low-level signals in the same thermal environment and surroundings, thus resulting in less thermal hysteresis and better turn-on characteristics. The size constraints and interconnection densities justify an integrated MEMS device when millions of display elements must be addressed and actuated at video frequencies. The same holds true when a handheld or medical device requires that thickness or width does not exceed system-level packaging limits.

The alternative is to create a larger signal, which often requires more silicon area, more power and lower impedances, and is implemented as a multichip solution. To gain enough information in the actual environment, a suite of sensors is added to the MEMS element and the signal is compensated in a remote location. The costs, size and complexity of interconnection scale nonlinearly, resulting in assemblies that are expensive and difficult to package and test.

Our experience with the single-chip gyro has shown that a signal at least two orders of magnitude less can be processed on-chip as compared to off-chip. The silicon sensor size typically must grow by factors of 10 to 1,000 in area to create equivalent signal-to-noise levels, which adds major design challenges for stress management and stability over time.

But integration is not the best solution for many MEMS devices. Small piezoresistive pressure sensors with low impedance have provided the MEMS industry with automotive and medical applications. High-temperature devices and chemical sensors take advantage of segregating the electronics from severe environments where silicon cannot operate. Costs and time may dictate that investments must wait until the system matures or the market can cover development costs.

Integrated MEMS products, even with a fully developed process in place, require a longer time than a nonintegrated version to go through each of the learning cycles. The tooling costs for masks can add up to $50,000 or more with each design iteration, taking several months to process with additional mask costs occurring at each change. Due to the close interaction between the MEMS design and the processing, it is difficult to confirm a viable process until several different designs have been built and the yields fully characterized. The dilemma of needing to concurrently optimize the process and the product design has no quick fix and often results in suboptimum "don't change anything" declarations from the customer.

Without large volumes that are predictable over several years, integration of MEMS just is not economically feasible. Saving $1 on 100,000 parts per year cannot offset spending $10 in development per part. This is further impacted by the uncertainty of customer relationships several years in the future. More subtle issues include the small additional cost of multichip packaging in a single package when economics do not justify the added costs of integration to a single chip in a smaller package.

Nonintegrated solutions may be the best technical and economic approach for a product. Larger silicon is sometimes needed to sense or interact with the environment. The simpler processing to create a mechanical structure with embedded strain gauges or a capacitive diaphragm configuration is far less expensive than an IC and MEMS process. RF modules and optical assemblies that meld the physical alignment, and the heterogeneous combination of different materials and processes, make a good point for some nonintegrated MEMS.

The drive to integrated MEMS has some compelling economic advantages when one considers the $3 accelerometer in widespread use today or the million-element MEMS DLP chip that makes possible a new era in displays.

Communication and connectivity have pushed MEMS devices to add complexity and functionality. The next generation of automotive sensors for safety systems is in development with a two-wire interface. This data and power over the same single-wire configuration creates a bidirectional bus with serialization, electronic diagnostics and calibration all performed in a standardized format. With the buying power of more than 100 million automotive safety devices per year, a new era in sensor interfaces is approaching. Coupling this with the wireless networks such as the Zigbee Alliance, the potential for MEMS devices that communicate along standardized protocols is being driven to larger volumes and more integration.

Processing options that build on deep reactive-ion etching make it possible to create MEMS structures after significant standard IC processing is completed. When this is combined with silicon-on-insulator wafers that have buried oxide, the prospect for integrated MEMS becomes more feasible with IC foundry processes. The buried oxide makes it possible to clear or release a MEMS device that has been surrounded by submicron CMOS processing on a single die.

The compelling argument for integration occurs when functionality must increase, price must be significantly reduced, and quality and reliability cannot be compromised. Then integration always wins.

Bob Sulouff is with the Micromachined Products Division at Analog Devices Inc. (Cambridge, Mass.).

See related chart