those of press-packed GTOs. Although first. IGBT-press-pack solutions have been presented to the market, IGBT-presspacking remains difficult and expensive.
IGBT POWER SEMICONDUCTOR RELIABILITY ANALYSIS FOR TRACTION APPLICATION Peter Jacob, Marcel Held, Paolo Scacco Swiss Federal Institute of Technology Zurich
with its usually short switching times (see below), these shearing forces are most effective to those structures with a short time constant - especially the wire-bond pads. Additionally, solder voids can occur in the die attach or in the DCB-heatsink solder layer. They increase the thermal resistance locally, raising the temperature swing at this point. In case of big solder voids, the points of lifted wire bonds corresponded very well with the voids; the latter could be non-destructively localized by using X-ray analysis. A similar study has been made in 1993 by INRETS in France  and ETH Zurich. Other failure mechanisms have been observed, too, like bondwire-melting and reconstruction of the chip metallization. In accelerated tests, meanwhile the first IGBT modules have passed the lomillion power cycle limit at a low ATj range of 30K. After module opening, corrosion to bond wires was found, which,however might come from humidity caused by deep cooling water temperature. Fig.l (at the end of the paper) introduces into important failure mechanisms.
IGBT (=Insulated Gate Bipolar Transistor) power semiconductor modules become of importance for traction applications. However, the reliability of these power modules is still a concern, compared to those of press-packed GTOs. Although first IGBT-press-pack solutions have been presented to the market, IGBT-presspacking remains difficult and expensive. Besides this, customers prefer module technology solutions, because of their economical advantage and easiness of device mounting. Reliability tests have been developed and modules have been analyzed to judge critical reliability points and to show ways for improvement.
Reliability Problems of Power Modules: Together with ABB Industrie AG, ETHZ started a series of analysis about the reliability of IGBT power modules of different manufacturers in 1993. Basically, it was found that within the active power cycling test, most failures were observed. After between 80.000 and 200.000 power cycles with temperature swings of 70K, most of the modules failed by wire-bond lifting. The higher the temperature swing, the earlier failures could be observed. The packaging setup of these modules consists of different materials with different thermal expansion coefficients [13. This causes shearing forces. In the power cycling test
0-7803-2797-7/95/$4.000 1995 IEEE
Customer Requirements While only in few cases 1OO.OOO power cycles could be survived by non-derating operation of the modules, traction customers require power cycling-reliability up to lomillion power cycles: This calculation is based on a lifetime of 30 years for traction vehicles like for example a tramway car. This request is,however,
5th IPFA '95:Singapore
away from the presently measured power cycling results of IGBTs. On the other hand, test results have shown, that a lifetime of 10 million cycles can be achieved, if suitable derating operation conditions would be accepted by the customer. Additionally, physical analysis has shown groundrules to optimize packaging design to high power cycling capability.
with only low power load. Also this case would not simulate the real operation situation. Thermal measurements by means of an infrared camera have shown, that above a gate voltage of 10 volts, the thermal homogenity from chip-to-chip within the module package is reached. However, to keep a guardband, it was proposed to fix the lowest gate voltage at 13 instead of 10 volt. b) Temperature swing: The junction temperature swing ATj should be 70K (or specified to application) at nominal current. The junction temperature must not exceed 125C. The temperatures may be measured electrically or calculated by using data sheet Rth values, or, if available, laboratory measured values. Tests, which have been done with a constant temperature swing but with different fixings of either the Tj(l0W) or Tj(high) temperature point have shown, that in case of fixing the Tj(high) point (usually at 125C) the cycle-lifetime is significantly lower than at tests with the Tj(l0W) point fixed (usually at environment temperature). In order to get a realistic reliability estimation of the devices, the operation conditions in the real application must be carefully considered. c) tonlt o f - switching and number of power cycles: The on/ off- switching must be done externally, that means, the DUT-gate is constantly open. On- and off-times are subject to the test engineer. The lowest time allowed is each on/off time=O,5 seconds for the highly accelerated power cycling test. However, this test is only good to monitor the wire-bonding reliability. To analyze solder reliability, either much longer cycles or passive thermal cycling are necessary. The time ratio may also be used to adjust the temperature swing. Looking to other failure mechanisms, the tests with short times (up to 1 sec) did show significance not only to wire bonding failures, but also to bondwire melting and
Testing Methods Unfortunetaly, conditions for most IGBTmodule reliability tests have not been standardized, yet. An inquiry at different IGBT manufacturers and users resulted in very different testing methods. Temperature swing, on- and off- times, cooling conditions and gate voltages of power cycling tests were defined very different. This was the initial reason to found two European working groups consisting of IGBT-manufacturers, users in traction applications, scientific laboratories and vendors of module-packaging materials and -tools. One group develops reliability tests, while the other one deals with technological improvements of the modules to increase their reliability. Meanwhile, the tester group has agreed on an accelerated power cycling test which runs as following: a) Power application to the module: The nominal current should be applied to the module. To control the power, the gate voltage may be adjusted in-between 13-15 volts. Best recommended value is 15V.This is to ensure, that the gate channel is open and no power is implemented to the module by a gate-controlled C-E-voltage value. This is important, because threshold voltage differences of single chips within the module might cause different CE-voltageand thus-power values. In this case, the thermal distribution within the module would become inhomogenious, caused by chips running under high power and such
5th ZPFA '95: Singapore
chip metallization reconstruction. In general, it is of course desirable to apply on/ off-times as long as possible in order to be relevant to most of possible failure mechanisms. However, this is mainly a problem of testing time. Very long tests with more than 1 Mio. power cycles to be expected, usually operate at derated conditions - in this case, the engineer would not expect unknown additional failure mechanisms caused by fast cycling. However,we must keep in mind that we observed unusual (probably testing-caused-) corrossion on wires of our 10-million cycles modules, which were still functional when we fmished the test. d) Cooling: The cooling unit must exceed the module's baseplate outlines in minimum 20" on each side. The surface planarity and roughness of the cooling unit must both be better than 10pm. When mounting the module's heatsink to the cooling unit, thermal conductive paste must be used and the manufacturer's mounting instructions (screw torques etc.) must be obeyed. The cooling liquid is water at room temperature. e) Failure Criteria; Before starting the test, an initial electrical characterization of the DUT-module must be done, including a threshold voltage characteristics measurement and a gate leakage measurement at 20 volts. The DUT is regarded to be defect, if one or the following criteria can be observed: 1.Vce deviation from the starting value of more than 20% divided by the #of chips within theDUT-module (if it is not known, it can be found by X-ray analysis). This measurement must been done at a fixed gate voltage of 15volt at nominal current and the module mounted as described in d). We have to add, that this criteria has been not very suitable in practice, and, thus it has been modified: Our tests have shown sometimes very different Vce vs.#of power cycles characteristics; if the (most usual)
bondpad lifting is the reason of failure, the characteristics shows no VCE increase for a long time and then, suddenly exponential increase shows the "chain reaction" of lifting bondpads. If, however, this mechanism is overlayed by chip metallization reconstruction, Vce shows a linear increase from the very beginning of the test. Also, the chip-number-dependend failure criteria generates disadvantages to modules with many chips inside, since a small deviation would be seen already as a failure in this case. In this case it could not be said, whether the increase is really caused by one defective chip or only a small voltage increase which might have uncritical reasons like accuracy of measurement. Facing these arguments, Vce increase will be observed at 5 , 10 and 15 % of the starting value in order to obtain the characteristics which might even allow to conclude to the failure mechanism type. After collecting some experience, the final criterium will be fixed. 2.Threshold voltage shift of more than 20% compared to the starting value. To do this measurement, the cycle routine must be interrupted at 5000, 10.000, 20.000..,etc. 100.000, than 500.000 cycles. The measurement must be done without a gate resistor. Characteristics comparison should be taken. 3 . I n c r e a s e o f t e le-e current to more than 1p.A at a gate voltage of 20 volts. This test will be taken at the same times like the Vt measurement. All these measurements are done at 25C. j') How to performpower cycling reliability data information: The reliability data should be given in a double-logarithmic diagram, where the horizontal axis shows the #of power cycles in-between lOexp4 and lOexp7. The vertical axis shows the temperature swing between 30 and 115C. The reliability curve displayed must show a proven measurement
5th IPFA '95: Singapore
point at its beginning and at its end; extrapolation is not allowed. g ) How to determine the nominal testing current ? Since the definition of the nominal current varies from manufacturer to manufacturer, it became difficult to perform comparable reliability tests. Therefore, the testing group defined a nominal testing current as following: In general, the typical output characteristics of the IGBT are given by the databook (or they can be measured easily), see fig. 2 below.
water cooling, T'usually is fixed to 45C, in case of air cooling or indirect water cooling, T' is 65C. Using these data, the solution equation is given by the type ax2+bx+c=0, Thus: Inom= - v C E O + ( v C E 0 2 + 4 ~ X r ( C E ) ) Ov5 2W.E) Practical expenence shows,that the nominal testing current is usually a little bit above the nominal current given in data books. This is due to the fact, that by reason of manufacturing, Rth has a certain deviation from module to module. To be on the safe side,IGBT manufacturers give "worst case"values, meaning to decrease the nominal current slightly below the physical values.
Applying these testing conditions, wire bonding reliability analysis could be made successfully within a reasonable time frame. While the tests used before needed times of months before reaching the 100.000 cycle point, this test performs this information within less than one week. A difficult item in reliability testing is the characterization of solder fatigue in area solder joints, like the DCB fixing on the baseplate. By the time, voids and cracks grow, increasing locally the thermal resistivity. If big voids are present (initially, caused by the solder process), they can be detected by means of Xray-analysis. However, if the module setup is too thick (for example in case of integrated water cooling unit) or if an Rth increase is caused by delamination (cracks) in the solder layer, X-ray analysis won't be a suitable instrument for analysis. A lot of attempts have been made by ultrasonic microscopy without success. The problem of this method is, that, when coming from the back side, some baseplates are too thick to get a reasonable signal (>8mm Cu), and, if coming from the chip side, wire-bonding and other layers influence the signal in a strong way.
Fig.2:Typical output characteristics of an IGBT. The curve taken at a gate voltage of 15V and a temperature of 125C (junction) should be used. This output characteristics has a linear increasing part. The gradient of this part can be geometrically determined (dV/dI) and gives the internal resistivity r(CE) .To determine the module-internal power loss, the voltage VCEO,which is the crosspoint between the linear extrapolation of the curve gradient in the linear part to the zerocurrent-line (x-axis), must be considered also. From these considerations, the following equation would result for the power loss P:
P=(VcEoxInom)+(rcE x Inom2)
On the other hand, the power loss can be determined also by using the thermal resistance Rth (either measured or taken from the data book) and the temperature swing ATSTvjmax-T'. Further: P=AT/Rth. Usually, Tvjmax is 125C. In case of direct
5th IPFA '95:Singapore
heats the chip. This additional heating can generate more or less chip metallization reconstruction. In severe cases cracks grow within the metallization layer, generating serious damage and even may disconnect contact holes at the chip metallization level [6,7]. In order to avoid this effect, a thoroughly covering with nitride/polyimide passivation (soft silicone is not suitable against this) can be applied. Polyimid is a suitable material against this, and it will also fix the bondpad mechanically in a certain way, even if it has been sheared already by thermal cracking. With respect to DCB material, AlN is of better thermal conductivity than A1203. However, A1N is the more expensive material and also additional material characteristics like elasticity, thermal extension and partial discharge behaviour must be considered. Another critical point is the large-scale soldering in-between the DCB and the heatsink-baseplate: If the DCB is not separated into small sections, the risk of large solder voids increases. Voids with a size of more than 510% of the silicon chip area become of serious influence to the thermal behaviour, and, if wire-bonds are located within such areas, they will fail at first. It is a self-accelerating failure mechanism: If one bondpad has sheared off, the other bonds must carry its current load. Thus, the temperature increases again, accelerating the failing of the remaining bondpads, and so on. Since big voids are (if present) located in the DCB center, the heating devices are best placed at the DCB border. Additional experiments were made together with the partners of the experimental group. These experiments dealt with the wire material, the wedge bonding tool shape, the wire-bonding parameters and bondpadpassivation techniques. Other experiment's goal is to find out early electrical and physical reliability indicators,
A suitable method is to detect the stressrelated change of hot spots on the module's active devices by means of infrared thermography. This needs a careful initial characterization by means of X-ray and infrared thermography. If changes are found, a final metallographic analysis can show the physical reason of degradation. This method can only be applied, if changes in the thermal homogenity caused by wirebond-liftoffs can be excluded.
Technological analysis and Experiments:
The reliability limiters are wire-bonding, reconstruction of metallization, wire melting and -corrosion and die attach-voids. The analysis of modules from different manufacturers could give important hints how to improve the module reliability without changing the basic technology. Speckle-interferometric- and infrared analysis have shown, that the time-constant of thermomechanical movement of the bondpad is less than half a second, and, that the bond heel suffers a heat concentration. Since the current load in the bond wires is up to 1OA for wire-thicknesses between 300 and 500pm (representing current densities up to 200 A/&), the heating contribution of the wire is of significant magnitude. Bondpad liftoffs never were observed on DCB copper conductive lines, but always on active silicon chips. On their surface, the thennal contribution of the junction and those of the wire add themselves resulting in critical thermomechanical shearing. The only cooling for the wire is given through the bondpads; the silicon gel in which the wires are imbedded, can be neglected. Thus, the wire-bond connections should be kept as short as possible. Infrared thermal analysis has shown, that below a certain specific current, the bondwire acts as a cooling for the chip. Above this value (which is the usual operation mode), it is viceversa: the wire
5th IPFA '95:Singapore
reasonable temperature swing (less than 70K) in the active power cycling test, the results of improved state-of-the-art modules reach quite close to the customer requirements. Technological improvements are ongoing continuously. As an example, the Low Temperature Joining Technique, which has been developed at Siemens and TU Braunschweig, should be mentioned . The first IGBT modules have been manufactured in this technology and are presently under test. In the future, the voltage ranges of IGBT modules will approach to those of GTOs. Therefore, the partial discharge test will become more and more important with special respect to ageing effects of insulator layers (DCB).
where already significant progress could be reached: The increase characteristics of VCE vs. #of power cycles allows to conclude on the main types of failures and even on the remaining lifetime. The results up to now look promising and show the way to highly reliable IGBTmodules for traction applications, based on the present technology without stressing the given economical frame for manufacturing.
Discussion The weak point is still the very long lifetime behaviour of the modules. No suitable accelerated test is still available to judge the 30-years-long-term behaviour of the joining techniques of the different materials within the module setup. We have done additional thermal, humidity, corrosive and vibration tests. Weak points could be identified (housing, wire material, encapsulation etc.) but were soon resolved and not too serious. Further efforts must be done to understand the failure mechanisms generated by our 10 million cycle tests. Although our work concentrates on packaging technology, silicon chip aspects must also be kept in mind. Finally, it should be mentioned here, that IGBT power modules are very advantageous devices to be applied in traction power converters. First of them have been already put into service of modern low-floor tramway trains. Their reliability can be estimated very good for different application conditions.
Acknowledgements: "Thank You" to all the companies and institutions attending to our tester standardization and experimental working groups for the excellent and fruitful cooperation: ABB (CH), ABB-Henschel @), AEG Westinghouse @), Aprova (CH), CEM (D), ETHZ CH), DuPont de Nemours @), Eupec (D), GEC Alsthom (F), INRETS (F), IXYS Semiconductors (D), Muller Feindraht/ AFW (CH), Orthodyne electronics (USA), Semikron Elektronik (D), Schindler Ascensori e Motori (CH), Siemens @), TU Braunschweig @).
References: [11 W.Wu,M.Held,P. Jacob; Reliability and Thermal Stress Aspects of Power IGBT Packages, ISHM Milano 1994 [Z] G.Coquery, Thermal cycling: first comparative tests on presspacked and direct bonding technology for GTOs and IGBTs, ESREF Bordeaux, Oct.. 1993 [31 P. Jacob ,M .Held,P. S cacco ,W. Wu ; Reliability Testing and Analysis of IGBT
Conclusion and outlook Reliability tests have been shown to monitor wire-bondpad shearing and other reliability limiters of IGBT power modules. The reliability can be improved by observing some basic packaging design rules, resulting from failure analysis. Using optimized IGBT-modules within a
5th IPFA '95: Singapore
Power Semiconductor Modules, ISTFA Los Angeles, Nov. 1994  ABB-Henschel data document; LESIT PZT report L30, ETH Zurich Feb.9,94 [SI W.Wu,M.Held,P.Jacob; Thermal Stress Related Die Bonding Failures in Power IGBT Modules (applied for publication at IEEE MCMC 95, Santa Cruz)  L. Yau, C. Hong, D. Crook;Passivation Material and Thickness Effects on the MTTF of A1-Si Metallization. 23rd IRPS, 1985  R.A. Schwarzer, D.Gerth; The Effect of Grain Orintation on the Relaxation of Thermomechanical Stress and Hillock Growth in Al-l%Si Conductor Layers on Silicon Substrates. Journ. of Electronic Materials, Vo1.22,No.6,1993  D. Gerth, D. Katzer, R.A. Schwarzer; The Influence of Local Thermomechanical Stress on Grain Growth in Thin Al-1%Si Layers. Phys. stat.so1. (a) 146,299(1994)  S.Klaka, R.Sittig, Reduction of Thermomechanical Stress by Applying a Low Temperature Joining Technique, ISPSD 94, Davos
cross section along the bondpad after power cycling: a crack has formed, interrupting the electrical and mechanical connection.
The cross section, cutting along the wire, shows voids and bubbles, caused by wire-intemal material
Fig. 1: Important IGBT reliability failure mechanisms: n after power cycling with high temperature swing: left half: passivated area, without defects, right half: nonnassivated."reconstruction".
Wire-bond liftoff after active power cycling test. The chip metallization is not hurted, wire metal remains on the chip surface
Thermography on a wire bond at different currents. At low current, the wire acts to cool the chip, with incrcasing currcnt, it becomes reverse.(by CEM)
5th ZPFA '95:Singapore