Technical Insight
Understanding imperfections in GaN HEMTs
Concerns over reliability testing and material quality are tempering the success of the GaN HEMT
RICHARD STEVENSON REPORTS
GaN has some great characteristics: It has a very high electric field strength, which allows the construction of devices operating at very high voltages; it has a high thermal conductivity, so chip overheating is not a big issue; and it has good carrier transport properties, which enable the construction of transistors operating at high frequencies.
Thanks to all these attributes, the GaN HEMT has a very promising future. It is tipped to wrestle market share from silicon in the power electronics industry, and in the RF arena it is expected to replace GaAs devices and vacuum tubes, due to its combination of robustness and high output power.
Success is not a given, however. Instead, there are still some weaknesses to address, which are related to material quality and reliability.
Many of these concerns over the capability of GaN HEMTs were highlighted at this year's CS Mantech conference, which was held in late May in Denver, CO. At this gathering Glen David Via from the Air Force Research Laboratory (AFRL) discussed the weaknesses associated with the existing approach for reliability testing; Feng Goa from MIT unveiled an electrochemical degradation mechanism in GaN HEMTs; Petra Specht from the University of California, Berkeley, exposed defective regions within the transistor; and Yutaka Tokuda from Aichi Institute of Technology in Japan reported the results of a comprehensive study into electron and hole traps in GaN epilayers grown on GaN, sapphire and SiC substrates.
Assessing reliability
Via argued that the accelerated lifetime tests applied to GaN HEMTs might fail to give a true picture of their reliability. That's because lifetime estimates are often obtained by taking data over relatively short time frames, and then performing extrapolations by several orders of magnitude. This means that lifetime derived by this approach could shift significantly if inaccurate estimates for channel temperature are employed, or there are errors in the temperature dependence of a material's thermal conductivity.
"My recommendations for increasing user confidence in lifetime tests are to conduct longer term, lower-temperature testing," says Via. "This will aid in reducing the risk of not uncovering a buried, low-activation energy failure mechanism."
He pointed out that another weakness associated with reliability testing is that the data reported is based on accelerated life tests of relatively small sample sizes. "In addition, population statistics are not commonly described, thus confidence bounds on extrapolated lifetimes cannot be determined."
What Via would like to see is a shift from studies that focus on the mean time-to-failure (MTTF) "“ that is, the time when 50 percent of devices have failed "“ to predictions of the time to first or one percent failure. "[These figures] will determine insertion readiness for mission critical systems."
At the meeting in Denver, Via presented a survey of industrially reported or published results for AlGaN/GaN HEMTs grown on SiC. He excluded devices formed on silicon from this study, because he only wanted to consider GaN HEMTs employed for mission critical and military applications. These types of applications require a MTTF of 100 years, which equates to a million hours, while commercial applications require a lifetime of just 20 years.
Via's study considered multiple-temperature DC and RF accelerated life tests with channel temperatures ranging from 250 °C to 400 °C, and drain voltages spanning 20 V to 65 V. Values for MTTF varied from less than 100 hours to more than 10,000 hours, and the vast majority of data predicted a MTTF in excess of a million hours for standard operating, which involves a channel temperature of 150 °C. "Given the available lifetime data, there should be little concern with reliability for commercial applications," argues Via.
To predict lifetimes, researchers employ Arrehenius plots, which include an activation energy "“ it is the slope of a line that is fitted to the MTTF.
"Ideally, the activation energy should correspond to documented failure mechanisms, as with silicon and gallium arsenide," says Via. However, at present, the GaN community is yet to agree on the dominant failure mechanisms associated with the activation energy. "The crack or pit that has been observed by many groups has been associated with activation energies ranging from 1.05 eV to over 2.2 eV."
To illustrate how many failure mechanisms have been proposed, Via presented the results of a survey of the literature by his colleague Eric Heller. This study uncovered reports of more than a dozen different explanations for GaN HEMT failure, including diffusion, defect percolation, surface barrier oxidation and trapping.
One interesting trend found by Via in his work is that the activation energy determined in lifetime tests is increasing, and it is now around 2.5 eV, compared with just 1.5 eV in 2007 (see Figure 1). He attributed part of this increase to improved material and processing, and he believes that this approach to lifetime testing is also leading to increases in activation energy values. " As the technology has matured, it has been more and more difficult to accelerate degradations in a reasonable amount of time, so harsher and harsher conditions have been applied." One major flaw resulting from this is that the induced failures are not representative of the "˜use' wear-out mechanism.
Figure 1: A survey of the literature by Glen David Via from the US Air Force Research Laboratory indicates a steady increase in the mean-time-to-failure of GaN HEMTs.
"There is also the complication that as GaN technology has become more rugged, life tests are not taken to failure," adds Via. Instead, engineers carry out the test for a fixed period, and times to failure are extrapolated to find a MTTF for the population. With this approach, lifetime estimates and values for the activation energy are then generated from an extrapolation of an extrapolation, increasing uncertainty.
Via has looked into possible correlations between activation energy and various device characteristics. The periphery of the device appears to have no impact on activation energy, which is lower when the gate length is shorter. Drain bias does not influence the activation energy, which is lower when using RF testing, rather than a DC approach, to perform accelerated tests.
Efforts at assessing the reliability of GaN-on-SiC through DC and RF accelerated tests and long-term operational tests will continue at the Sensors Directorate at AFRL. "The goal is to help the community build a better understanding of GaN degradation physics and increase the confidence in lifetime estimates through statistically relevant sampling of technology offerings," says Via.
Electrochemical reactions
At MIT, Gao and colleagues are claiming to have a comprehensive explanation for a key structural degradation mechanism in GaN HEMTs "“ the creation of electrical pits near the surface of the transistor.
"It was not clear why this degradation was only seen by some research groups and not others," says team leader Tomás Palacios. "We found that the difference between the samples of different groups is probably not only in the magnitude of the electric field but, especially, in the quality of the passivation."
Engineers from MIT found that pits occur when water reaches the III-N surface. This leads to the creation of an electrochemical cell, which forms at the gate edge where gate metal, III-N surface and passivation layer meet.
Uncovering this degradation mechanism involved tests under vacuum, ambient air and a water-saturated environment, using GaN HEMTs made both at MIT and at an industrial collaborator. These transistors consisted of a 3 nm-thick GaN cap, a 14 nm-thick AlGaN barrier, a 1 nm-thick AlN interlayer and a thick GaN buffer deposited on semi-insulating SiC.
Metal evaporation added a Pt/Au gate with a length of 250 nm, and a thick SiN layer deposited by plasma-enhanced CVD passivated the surface of the
device.
Testing began by taking two chips from the same wafer, each with five identical HEMTs. Both were subjected for 3000s to a high-drain off-state bias: a gate source voltage of -7 V, and a drain source voltage of 43 V (which is much lower than the breakdown voltage). One of these chips was held in ambient air, and the other in a 1 x 10-7 Torr vacuum.
HEMTs from both chips were scrutinised with a variety of microscopy techniques. Scanning electron microscopy and atomic force microscopy were used to study an exposed area formed from removal of the gate metal. This revealed that stressing in air, rather than vacuum, led to larger, more numerous pits (see Figure 2).
Figure 2: Experiments by Feng Gao and co-workers from MIT show that water is to blame for the creation of pits in GaN HEMTs. Under water-saturated argon gas ((a) & (b)), larger, more numerous pits form than in a dry argon environment ((c)&(d)).
These finding were confirmed by cross-sectional transmission electron microscopy (TEM) measurements on another pair of HEMTs that had not been subjected to etching of the passivation and gate metals.
The engineers discovered, via energy-dispersive X-ray analysis, that these pits had a high concentration of oxygen and a low concentration of aluminium and gallium. The two likely sources of oxygen were oxygen gas and water. To identify the culprit from these two suspects, the team took another two chips of five identical HEMTs and applied the same stress test as before. However, this time they placed these devices in either a water-saturated argon or dry argon environment. Many more pits were found in HEMTs stressed under water-saturated argon conditions, pointing the finger at water molecules as the cause of pitting.
To rule out other causes, the team considered other gases, such as oxygen, nitrogen and carbon dioxide, comparing results for a water-saturated and a dry environment. Pit sizes were always greater when water was present, and there was very little pitting when the only source of oxygen was dry oxygen gas or dry CO2.
The team argues that the root cause of the pitting is an electrochemical cell at the gate edge that causes anodic oxidation of the AlGaN layer. This reaction starts at the GaN surface and proceeds into the AlGaN barrier during electrical stress. The gate metal acting as the cathode provides electrons to the water at the SiN/AlGaN interface when the gate-to-drain diode is reverse biased, while trap-assisted inter-band tunnelling in the AlGaN barrier provides holes for the reaction.
MIT's findings may raise a few eyebrows, because a thick layer of SiN has been widely viewed as an excellent barrier to moisture. "However, there is nothing perfect," says Palacios, who points out that it actually takes a very small volume of water to penetrate through SiN and destroy a HEMT.
The obvious approach to reducing the formation of pits is to switch to a better water barrier. "Our group has used hydrophobic materials such as Teflon to delay these water-assisted electrochemical reactions and this largely improves the device reliability and lifetime," reveals Palacois.
He points out that this approach complements a well-trodden path for increasing device lifetimes: reducing electric field strengths via the introduction of field plates and other structures. Targets for his group are to understand how the electric field impacts the device, and how the degradation physically happens. "In collaboration with colleagues at MIT and elsewhere, we are now interested in studying if dislocations and other lattice defects play a role in accelerating device degradation."
Uncovering point defects
Another group looking at defects in GaN HEMTs is that involving Specht, plus colleagues at the Naval Research Lab and the Naval postgraduate School. Their efforts have focused on finding and identifying point defects with an ultra-high resolution, aberration-corrected TEM that can deliver local chemical analysis. Using this approach they separated effects coming from epilayer growth, device processing, transistor operation and the environment.
"The high-resolution imaging techniques we are using were developed for nano-structure and two dimensional material research, but have never been tried before on complete device structures," says Specht. "We have reason to expect that we can go beyond the current results available on GaN-based device reliability, finding and ambiguously identifying the sources of failure mechanisms."
Sample preparation involved a novel "˜all ion-mill' approach that begins with a high-energy, focused-ion beam lift-out process, followed by a low-energy
argon ion milling that enables local cleaning of the sample, while maintaining its structural integrity. Cutting involved a 30 kV gallium beam, while additional thinning and cleaning of the sample employed argon beams with energies of just 900 eV and 500 eV, respectively.
"The resulting clarity in our images is rarely seen anywhere else," claims Specht, who puts this down to the use of the world's best transmission electron aberration-corrected microscope, which has a resolution of 0.5 à… and is located at Lawrence Berkeley National Laboratory.
The team looked at a variety of HEMTs, which were all formed on silicon and featured buffer layers with a thickness of 1.2 µm, followed by an 800 nm-thick layer of GaN, a 17.5 nm layer of AlGaN and a 2 nm-thick GaN cap. The portfolio of devices under study included those with gates made from PtAu and NiAu, and those that were either fresh, had degraded after operation or failed during operation.
Imaging with the ultra-high-resolution TEM revealed two typical defects close to the gate edge: a gold "˜foot' area; and beside it, a crater-like "˜dip', which is much deeper and wider in devices with a PtAu gate (see Figure 3). Specht and co-workers believe that the relative location of these two extended defects "“ the foot and the dip "“ indicates that they are process related.
Figure 3: The ultra-high resolution, aberration-corrected transmission electron microscope at the Lawrence Berkeley National Laboratory can uncover the profile of gold near the foot of the gate of GaN HEMTs.
Energy dispersive spectroscopy unveiled gold contamination in the dip, which is believed to occur due to diffusion from the gate contact (see Figure 4). This technique also revealed a thinning of the AlGaN layer, with a larger, gold-containing amorphous layer forming at the surface.
Figure 4: Electron dispersion spectroscopy can identify the location of elements such as silicon, nitrogen, oxygen, aluminium, gallium, platinum, and gold. This measurement technique revealed: the absence of a GaN cover layer; the formation of AlOx at the surface; a decrease in the nitrogen profile, which occurs with increasing oxygen content; and the accumulation of gold at the surface of the "˜dip' area.
One of the most surprising findings of the study was the absence of the 2 nm-thick GaN cap in all structures apart from the epilayer.
"It is still unclear to me how uncharacteristic this is for the initial wafer," admits Specht. "The piece I got may have also come from the edge of the wafer, and may have had some inhomogeneous deposition there." It is also possible that an initial cleaning step applied to this piece, prior to device processing, could have led to the loss of the GaN cap. "We are clarifying those issues right now," says Specht.
Cataloguing traps
Meanwhile, Tokuda is continuing his comprehensive study of electron and hole traps in n-type GaN grown on sapphire, GaN and SiC. At CS Mantech he revealed that the combination of deep-level transient spectroscopy and minority carrier transient spectroscopy on Schottky diodes, p+n diodes and MOS structures had uncovered nine electron traps with energies between 0.24 eV and 1.2 eV, and five hole traps with energies ranging from 0.25 eV to 1.8 eV (see Figure 5).
Figure 5: Deep-level transient spectroscopy can reveal electron traps in GaN layers grown on a variety of substrates.
Of this total of 14 traps, three are commonly observed in MOCVD grown GaN, regardless of the substrate: electron traps at energies of 0.24-0.26 eV and 0.57-0.61 eV, and a hole trap with an energy of 0.86-0.88 eV. According to Tokuda, the lower energy electron trap may be due to dislocations; while its higher-energy cousin may depend on growth conditions, such as temperature and pressure; and the hole trap may depend on growth conditions.
"We are [now] studying the distribution of trap concentrations on each two-inch and three-inch MOCVD n-GaN wafer, to understand the relationship between the trap concentration and the performance of fabricated devices," says Tokuda.
This effort, plus those by other researchers that are studying the weaknesses in GaN HEMTs, will help to address the flaws of this device and ultimately drive its commercial success.
-Note that the views, opinions, and/or findings associated with Glen David Via should not be interpreted as representing the official views or policies, either expressed or implied, of the Air Force Research Laboratory or the Department of Defence.
