Typically, Nanomix sensors are based on CNT network FETs produced on 6-in. silicon wafers using the chemical vapor deposition process and certain proprietary methods. In some cases, flexible substrate and spray coating are used.

The CNT network is coated with a functional layer that interacts with the chemical or biological analyte of interest. Interactions between the functional layer and the analyte result in a measurable change in the electrical characteristics of the FET (Figure 1).

Figure 1. Electrical Diagram of a CNT-Network FET in a Gas Sensor

The nanometer-scale diameter of a CNT and its chemical and electrical properties allow ultrasensitive detection. The resulting sensors are small in size and low in power consumption.

During production, testing is done at the wafer level to evaluate the basic characteristics of the FETs. After the FETs are functionalized, the packaged devices are exposed to target analytes and their detection characteristics analyzed.

Wafer-Level Testing
In wafer-level testing, I_d-V_d and I_d-V_g measurements are taken at various stages of production. The tests are conducted on actual devices and special test structures. The test equipment includes a semi-automatic prober, Keithley Model 2400 SourceMeters, a Model 7002 Switch Matrix with 7011C Multiplexer Cards, and a PC that provides sequence control.

Testing takes place on a wafer prober without exposure to the target analyte. Analysis of the test results, such as gradients of certain characteristics on a wafer map, may be used for fine-tuning various stages of the development process. A QC test regimen is used during production for wafer disposition and die selection.

Typically, FET characteristics such as G_min and G_max are calculated from the raw I_d-V_g data. The raw data and the calculated characteristics are uploaded to a production database. The engineers then can generate histograms and interactive wafer maps from the database using proprietary software, inspect the maps for uniformity, and compare different types of devices that are present on the wafer (Figure 2). If necessary, statistical analysis is performed on multiple wafers or within a wafer.

Sensor Assembly Testing
Testing of packaged CNT-network FETs and chemical sensors using those FETs may take place inside environmental chambers. Usually the analyte gas mixture is supplied by one of the custom gas delivery systems (GDS) developed at Nanomix through special manifolds attached to the test fixture boards.

The DUTs have various package configurations. The number of DUTs can vary from one to more than 100. After being loaded onto appropriate test fixtures, the devices are exposed to analytes and changing environmental conditions during testing. This also may require monitoring with additional reference sensors.

Depending on the tests being run, DUTs may require uninterrupted bias voltage typically between 5 mV and 500 mV. Device conductances can range from 1 microsiemens (�S) to 100 millisiemens (mS) as determined by the design and fabrication methods.

In development test scenarios, the main objectives are sensor characterization and direct comparison of different fabrication methods through statistical analysis of measured parameters. In both development and production, a major test concern is high-throughput measurements of multiple devices over time.

To streamline testing and analysis, the DUTs are arranged in test groups, a technique known as blocking in experiment design. A test group may be composed of nominally identical devices, or conversely, some devices may be intentionally different.

The device identification information such as wafer, die location, package pinout, or functionalization type is entered only once and used for all experiments with the test group. All devices in a group are exposed to the same testing and storage conditions and the same sequence of tests, which reduces the possibility of mistaking unknown factors for intrinsic differences between devices.

Original Test System Design
Initially, the test system was composed of custom-designed DUT fixtures, onboard multiplexers, and current preamplifiers, all of which were interconnected with ADC and DAC boards. Test control was accomplished with a PC running LabVIEW from National Instruments.

However, test fixtures were large, complicated, difficult to maintain and troubleshoot, and supported only 40-pin DIP packages and could not operate reliably at higher than room temperatures. When multiplexing the measurements, maintaining a constant bias voltage on all devices was impossible.

Moreover, measurements had a limited range and insufficient resolution, with cycle times that were considered excessive. Test control also was becoming quite difficult due to test complexity and the need to quickly accommodate changing requirements. As a consequence, making changes on the fly required maintaining multiple software versions.

Improving Test Parameters
To overcome these limitations, we developed an automated test system called Zephyr. One of the main challenges was the flexibility required to handle multiple package types and test conditions.

Simple passive test fixtures were built to house several different types of DUT packages. These are connected to the test system over 1.5-m twisted-pair DB25 cables, and the fixtures can be placed in the environmental chamber. With this hardware, DUTs can be heated or cooled individually.

An important benefit of the Zephyr system is its flexibility. It is designed to work with different measurement equipment configurations as tests demand. One typical application uses a Keithley switch matrix with two multiplexer cards, one Model 2400 SourceMeter, one Model 2602 SourceMeter System, and a custom Trigger Link adaptor. More complex tests or those with a larger number of DUTs may use additional switch matrices, switch cards, and source measure units (SMUs).

The test executive software is written in Java and is platform independent. The current version includes only Windows drivers for serial ports and GPIB, but adding new drivers is trivial. Instruments from multiple vendors are supported. The tests are programmed using a specialized script language, which is based on XML.

Using the supplied schema, script files can be written in an editor with autocompletion and syntax checking. The script files also can be source controlled, which makes it possible to track changes or go back to earlier versions if necessary.

The flexibility of Zephyr software is due to loose coupling between the following features:

� The test script determines the sequence and timing of all switch and measure operations, which encompass both DUT and reference measurements as well as heater control and certain GDS control functions. This approach is known as domain-specific programming language.

� Test-script settings such as device addresses, voltage levels, test duration, and measurement frequency can be modified for every test run and installation without changes to the script.

� Instrument control provides low-level implementation of communications in-
terfaces shared by all scripts. Instruments from multiple vendors are supported. All instrument adapters use the same application programming interface (API), making instruments generally interchangeable with only minor changes to the script. (One instrument adapter collects target values from the GDS).

� The software facilitates DUT and test-group identification. The application can use test-group data to convert generic measurement names specified in the script to specific device names.

� The script does not handle data output explicitly. Instead, the application sends measurement data to every active output component, which stores or displays relevant data in a specific format. These components also use a certain API and can be controlled by the user.

Some tests take advantage of the Test Script Processor (TSP�) built into the Keithley 2600 Series system. The TSP scripts, which are based on a widely used programming language, can range from just a few commands to a large set defining sophisticated test functions. These scripts are quite different from the Zephyr scripts. However, if necessary, it is possible to upload and run a TSP script on the Keithley 2600 System using Zephyr software.

The most recent feature added to the software is the capability to measure the I-V characteristics of the devices without increasing the level of Zephyr test script complexity (Figure 3). This is achieved by adding a virtual instrument or instrument adapter in Zephyr terminology that returns a special type of measurement result, which is handled by the appropriate output components.

Behind the scenes, the adapter uploads a custom TSP script to a Keithley 2602 at the beginning of a test. For every measurement request from the Zephyr script, the adapter calls a function defined in the uploaded script and transforms the returned data for further processing by the output components. Accordingly, a new type of measurement can take extensive advantage of hardware capabilities and be integrated seamlessly into the system.

About 10 Zephyr scripts currently are in use for various purposes. For example, with typical settings, one of the scripts used for DC measures 16 devices every 500 ms repeatedly for several hours. The test duration or number of cycles is set by the operator and depends on the specific task.

By taking advantage of the switch matrix and SMU capabilities, the system has greatly increased test throughput. In certain applications that were possible with the old system, the cycle time was decreased by a factor of three, and low-conductance accuracy improved by at least two orders of magnitude and resolution by at least one order of magnitude.

In reference device tests of the current system with the Keithley Model 2602 set at one power line cycle, the resolution limited by noise was better than 250 pA or 5 nanosiemens (nS) at 50 mV. The accuracy for low-conductance devices was better than these values. For low-resistance devices, the accuracy of approximately 1.5 Ω was determined by cables and relay contact resistances.

These characteristics, which are sufficient in this case, can be improved if necessary. For example, low-resistance accuracy can be greatly enhanced by using four-wire connections, which would require a special fixture and twice as many cables.

Better accuracy also can be achieved by adding a short to the test fixture and programming automatic correction in the script. Shielding fixture connection cables and the fixture itself could further decrease the noise. Triax connections may be used for very low-conductance devices.

Conclusion
Zephyr now has been in use for almost two years, and we have found it to be a highly scalable and flexible test system. Currently, there are multiple installations in various configurations. Support for wafer-level tests and autoprober control is planned for future versions. This will replace older software, allow an upgrade of the autoprober measurement system from the Keithley Model 2400 to Model 2602, and further decrease test cycle time.

In February, the Nanomix Quality Management System received ISO 9001:2000 and ISO 13485:2003 certification. Relying on simple test fixture design, standard measurement equipment, source-controlled measurement software, and scripts contributed to the validation of the Zephyr Measurement System and overall certification of the company.

Although the system is used extensively, it still is a work in progress. The Zephyr software is released as an open-source project at http://zephyr.sourceforge.net/.

About the Author
Sergei Skarupo is the senior software engineer and the lead designer of Zephyr software at Nanomix. As a software engineer, he has been involved in development and support of various medical, Internet, analysis, and lab measurement applications. He studied electrical engineering at the Kiev Polytechnic Institute. Nanomix, 5980 Horton St., Suite 600, Emeryville, CA 94608, 510-428-5337, e-mail: sskarupo@nano.com