1 Hybrid Extraction Method for Determining Circuit Elements of AlGaN/GaN HFETs By QIAN FAN, JACOB H. LEACH, Student Member, IEEE AND HADIS MORKOç ABSTRACT: The developments in AlGaN/GaN heterojunction field effect transistors (HFETs) are beginning to allow harnessing of the great potential of this technology in high power RF applications. However, the integration of HFETs into a circuit environment requires accurate small and large signal modeling of the device operating under various biasing conditions. The conventional small signal equivalent circuit modeling methods consist of “cold” measurements for extracting parasitic elements, and on-bias measurements in determining the intrinsic device circuit elements. Also, the optimization routines are often explored directly using the “hot” measurement data to minimize errors. In this paper, an 18-element small signal equivalent circuit model for AlGaN/GaN HFETs is proposed and implemented, and contrasted to various de-embedding methods. Among the methods treated, the hot-FET optimization extraction based on the parasitic capacitances obtained from cold measurements leads to the smallest error between the simulated S-parameters and the measured ones at all bias points employed, with an average error of about 5%. This hybrid extraction algorithm is strengthened by imposing constraints to avoid any non-physical convergence. We note that the extrinsic parasitics determined by this method differ considerably from the values obtained by cold-FET measurements, which implies that the assumption on the bias independency for extrinsic parameters in the latter method might be questionable for AlGaN/GaN HFETs. KEYWORDS: Gallium Nitride, Heterojunction Field Effect Transistors, Equivalent Circuits, Small Signal, Parasitic I. INTRODUCTION T HE AlGaN/GaN HFETs have already demonstrated many advantages over GaAs-based devices for microwave power applications due to the unique properties of GaN-based materials such as high breakdown voltage, high carrier concentration and along with reasonable carrier mobility, and high thermal conductivity as described in [1]. A typical AlGaN/GaN HFET device structure along with its conduction band diagram is shown in Fig. 1. Because of the strong spontaneous and piezoelectric polarization of GaN and its ternary, a high density two dimensional electron gas (2DEG) is induced at the heterojunction interface and confined in the triangular potential well due to the large band discontinuity. The high electron velocity, which is estimated above 107 cm/s, bodes well for high frequency operation, and the large breakdown voltage allows for high power applications to be exploited. In this endeavor, the GaN-based HFETs exhibit great potential in RF power and switching applications. With the rapid development of semiconductor epitaxy, device processing, and system integration, an industrial supply chain for GaN power devices is emerging. Fig. 1 The schematic of AlGaN/GaN HFETs structure, conduction band diagram and electron distribution. It should be noted, however, that any successful system level design of power amplifiers, oscillators, or mixers requires accurate empirical device modeling, which must provide a circuit level description of a device's nonlinear properties. In this vein, small signal equivalent circuit modeling plays an important role in gaining the much needed insight into device nonlinearity at least in the static sense when carried out at different bias conditions. Clearly, the validity of device modeling is dependent on the accuracy of the small signal extraction routine employed. Standard small signal extraction methods [2], [3] are well-established for GaAs HFETs or SiC MESFETs. Basically, the equivalent circuit can be divided into the intrinsic part, for which the values are a function of bias, and the extrinsic part that has no dependence on the bias conditions. Usually each element is obtained by fitting the scattering parameters measured directly from the device. Accurate measurement of S-parameter is imperative for determining the equivalent circuit parameters realistically which in turn paves the way for the extracted values to have any physical significance. Although not without controversy, a good step in this direction is “cold” measurements under channel pinch-off biasing conditions (VDS = 0V, VGS ≤ 0), at which the intrinsic circuit part can be simplified, is crucial for extracting the bias-independent extrinsic parameters such as parasitic resistances or capacitances. With the parasitic elements determined, confidence in determining the other parameters from s-parameter can be enhanced. The above mentioned extraction methods have been applied to GaN based HFETs [4], [5]. The most comprehensive model for a GaN HFET is proposed by Jamdal and Kompa in [6] and consists of 22 distributed elements, which is reliable, general, and applicable for large signal case as well. The circuit model for GaN HFETs is often more complicated than traditional 2 FETs in order to account for behavior somewhat unique to GaN such as the gate leakage current, self-heating, or defect-induced dispersion effects [7], [8]. The validity of the extraction of the bias-independent parasitic elements using the cold-FET methodology on GaN HFETs is still under some question [9], perhaps due to the relatively high contact resistance for GaN ohmic metallization [10]. Some hot-FET extraction methods have been reported in [11], [12], which attempt to optimize and calculate the equivalent circuit elements including both the intrinsic and extrinsic ones directly from data measured at the operation bias point. Full optimization procedures, in which the optimization algorithm used determines each and every element, are typically reported to provide very high agreement between simulation and measurement. However, the problems of non-physical or initial value dependent convergence shown in [13] make these methods less desirable, particularly if verification over a wider bias and scaling range is desired. In this paper, we propose and implement an 18-element small-signal model for GaN HFETs, and consider various extraction methods. A hybrid extraction technique combining a cold measurement together with an optimization routine is developed which leads to a better agreement between the measured and simulated S-parameters compared with conventional cold measurements. It also provides better accuracy over the entire range of biasing conditions. II. DEVICE PROCESSING AND CHARACTERIZATION The Al0.25Ga0.75N/GaN heterojunction FET structure used in this investigation was grown on c-sapphire by low-pressure metal-organic chemical vapor deposition (MOCVD). First a 300 nm semi-insulating AlN buffer layer was deposited at 1050 ºC, followed by a 3 µm thick un-doped GaN layer and a thin AlN spacer with thickness ~1-2 nm. Next the 20 nm Al0.25Ga0.75N barrier layer was grown, intentionally doped with Si to 8×1017 cm-3, which does not really add much to the 2DEG next to the much larger polarization induced charge, followed by a 2 nm un-doped GaN cap layer. The temperature dependent transport properties from this structure are shown in Fig 2, as determined by Van der Pauw patterned Hall measurement. The room temperature electron sheet concentration is 1.10×1013 cm-2, with the mobility of 1200 cm2/Vs. The HFET devices were fabricated by first depositing the source/drain ohmic contacts using Ti/Al/Ni/Au and rapid thermal annealing (RTA) at 800 ºC for 1 minute in N2 ambient. Mesa isolation was then carried out by dry-etching in a parallel plate RIE system using BCl3 plasma. Finally, 2× 80 μm wide gates with a gate length of 1 μm were defined and Ni/Au was deposited as the gate metal. The DC performance of the devices was examined by a Keithley parameter analyzer, as shown in the inset of Fig 2. The appropriate quiescent point (Q-point) was determined to be between VDS = 5 ~ 7 V, VGS = -2 ~ -4 V. The on-wafer S-parameters of various devices from the same wafer were measured by an HP (Agilent at the moment) 8510B network analyzer with the highest frequency available being 20 GHz. The cold pinch-off measurement was carried out at VDS = 0 V, VGS = -7 V in order to subtract the pad parasitic capacitances. 2.4x10 13 500 V GS 400 4 1.2x10 4 =0~-5V 300 2x1013 1.5x10 200 9000 100 1.6x10 13 0 0 2 4 6 8 V (V) DS 1.2x10 13 8x10 12 10 6000 3000 0 50 100 150 200 250 300 0 350 Temperature (K) Fig. 2 The temperature dependent transport properties of the AlGaN/AlN/GaN HFET structure. The insert shows the typical IV (DC) characteristics of the HFETs. III. EQUIVALENT CIRCUITS As a three-terminal device, the extrinsic circuit for the FETs must include three parasitic resistances and inductances associated with each contact and wire connection. In principle, these elements should be as small as possible, although for wide band gap semiconductors the Ohmic contacts still require some attention. Regarding the capacitances, there are often two sets, namely the parasitic capacitances introduced by the pad connection or probe contacts described as Cgsp, Cdgp and Cdsp; and the those accounting for the inter-electrode capacitances such as Cgsi, Cdgi Cdsi. They are all implemented in [6], and the de-embedding procedure (an optimization-based routine) requires the initial values estimated from empirical assumptions. In order to reduce the complexity of the modeling procedure, the inter-electrode capacitances are often neglected by many authors [14], [15] since these values are always relatively small. In our effort, we propose an equivalent circuit for the GaN HFET as shown in Fig 3. The part outside the dashed box is the extrinsic one, in which the Cdgp is also neglected in favor of the simplicity of optimization routines as described below. Rfdg G Lg Rd Rg Cdg Rdg V Cgs Rfgs Ids Ld D Cds Rds Ri Cdsp Cgs Rs Ls S Ids=gmVie-jωτ S 3 Fig. 3 The proposed 18-elements small signal equivalent circuit for a GaN-based HFET, within the dashed box is the intrinsic part. For the intrinsic circuit model, the consistency between GaN HFETs and GaAs- or SiC-based power transistors is obvious. The traditional model often consists of 8 elements with explicit physical significance. For example, the charging resistance Ri represents the electron recombination with donors in the depletion region under the gate; and the transit time τ is the time it takes for the depletion region to respond to the change in the gate signal. But for GaN HFETs, the scenario is little more complicated. The differential resistances introduced in [15], Rfdg and Rfgs, are necessary for characterizing the current conduction through the gate diode. Reference [14] even considered the possible existence of time delays in the output conductance element (Rds). Addition of uncertainties onto the intrinsic circuit, however, will overly complicate the system of equations. To circumvent the ensuing complexity, we will examine a 10-element intrinsic circuit model similar to [6], which is believed to have sufficient accuracy and also applicable for large signal analysis. IV. COLD-FET EXTRACTION The key in the small signal de-embedding methodology lies in the determination of extrinsic circuit elements. Therefore, most efforts have concentrated on cold-FET measurements (VDS = 0) at near-zero drain current conditions to signify the parasitic parameters. A. Parasitic Capacitances The pad capacitances can be extracted under a pinched-off “cold-FET” condition (VDS = 0, VGS <<0) at low frequencies (in the MHz range) so that the influence of inductances can be minimized. Under the pinch-off bias conditions, the channel conductivity is negligible and the S-parameters measured exhibit capacitive properties. Consequently, the equivalent circuit complexity can be reduced as shown in the inset to Fig 4 if one considers only the imaginary part of the Y-parameters, [3], with the assumption of a symmetric gate depletion region. It then leads to the following relationships: (1) Im( Y11 ) = ω (C gsp + 2 ⋅ C b ) Im( Y12 ) = Im( Y21 ) = −ω C b (2) Im( Y22 ) = ω (C dsp + C b ) (3) where Cb is all of the remaining capacitances of the circuit. Cb 2.0x10 -3 C Cb C gsp dsp -Im(Y ) 12 1.5x10 -3 1.0x10 -3 5.0x10 -4 Im(Y )+Im(Y ) 22 12 Im(Y )+2Im(Y ) 11 0.0 0 9 2x10 9 4x10 ω(rad/s) 12 9 6x10 9 8x10 Fig. 4 Reduced equivalent circuit for pinch-off Cold-FET condition, and the parasitic pad capacitances determined by linear regression at VDS = 0 V, VGS = -7 V, frequencies from 150 MHz to 1.0 GHz. The pad capacitances extracted from a typical device under different bias conditions by calculating the slope of the imaginary Y-parameters linearly fitted over frequency are listed in Table 1. Higher applied biases result in stable values of these parasitics. As suggested also in [16], a high gate bias is desired to suppress the differential resistance and obtain the values for parasitics accurately. But a high gate bias would also introduce surface state-assisted tunneling charge and ensuing current flow, and subsequent gate breakdown problems, which, as we observed, fail to produce reliable capacitance values for some devices. Table 1 Gate Bias (V) Cgsp (pF) Cdsp (pF) -5 0.095 0.102 -6 0.081 0.054 -7 0.069 0.062 -8 0.063 0.061 -9 0.067 0.057 The extracted pad capacitances under different pinch-off bias. B. Parasitic Inductances and Resistances The parasitic inductances and resistances are generally measured under cold-FET conditions with large forward gate bias (VDS = 0, VGS >>0) in order to reduce the depletion capacitance to a large extent [17]. However, it has been pointed out by many groups that this bias condition is not applicable to GaN HFETs. Because of the existence of gate differential resistances, it often requires very high forward gate bias to eliminate the channel capacitive, which is usually accompanied by the unrecoverable damage onto the gate. Therefore, for GaN HFETs, the prevalent methods used to extract these inductances and resistances are often conducted at zero or small negative gate bias conditions. Another advantage for these bias conditions is the ability to simplify the circuit model to a symmetry form in favor of extraction. Under such bias conditions, after de-embedding the pad 4 capacitances measured from Section IV A, the rest of the device is modeled with equivalent circuit shown in Fig 5. 5.0x1011 G Lg Rg ΔZg Rd Cd ΔZd Cg Ld D ωIm(Z )-ωIm(Z ) 22 12 0.0 Cs ωIm(Z 12 ) ΔZs -5.0x1011 Rs ωIm(Z 11 )-ωIm(Z 12 ) -1.0x1012 Ls S S (a) 12 Fig. 5 Reduced equivalent circuit for low gate bias Cold-FET condition, after removing the pad capacitances. The residual intrinsic impedance terms (ΔZi=g,s,d) in Fig 5 are always treated as channel resistances. As such they have no influence on the imaginary part of the Z-parameters, from which the parasitic inductances are determined by linearly fitting Im(ωZij) over ω2: 1 1 (4) + Im( ω Z ) = ω 2 ( L + L ) − ( ) 11 s g Im( ω Z 22 ) = ω 2 ( L s + Ld ) − ( Im( ω Z 12 ) = ω 2 L s − 1 Cs Cs 1.0x1022 1.5x1022 1 1 + ) Cs Cd (5) (6) (8) (9) Re( Z 22 ) = R s + R d + R ch In order to determine four unknowns out of three equations, various bias points need to be measured to interpolate Rs+Rd, which was beautifully done in [4]. In our method, we neglected the channel resistance at zero gate bias, as suggested in [6], making the determination of parasitic resistances much easier. On the devices from which we extracted pad capacitances, the parasitic inductances and resistances were calculated over the high frequency range from 10 GHz to 20 GHz to minimize the error from residual capacitive terms. Fig 6 shows the linear fit of ωRe(Z) vs. ω, and ωIm(Z) vs. ω2 measured at zero gate bias. If we neglect the residual terms, then the resulting parasitics are: Rs = 4.45Ω, Rd = 5.47Ω, Rg = 3.69Ω, Ls = 6.82pH, Ld = 36.99pH, Lg = 52.19pH. As can be noted in Fig 6, the curve fittings for gate parasitics show a larger spread. They are also more sensitive to the frequency range in which the fitting is performed, similar to what was observed in [21]. Therefore, the cold measurement method provides a better accuracy and thus more confidence in the source and drain parasitic resistances and inductances than those of the gate. 2.0x1022 ω (rad /s ) 2 2 2 12 1x10 ωRe(Z12 ) 11 8x10 ωRe(Z22 )-ωRe(Z12 ) Cg The extraction of parasitic resistances, on the other hand, depends on ΔZi which is expressed in various ways in different reports. For example, from zero to a low negative gate bias range, it follows: (7) Re( Z 11 ) = R s + R g Re( Z 12 ) = R s + R ch / 2 -1.5x10 5.0x1021 11 6x10 11 4x10 ωRe(Z11 )-ωRe(Z12 ) (b) 11 2x10 10 8.0x10 1.0x10 11 11 1.2x10 1.4x10 11 ω(rad/s) Fig. 6 (a) Parasitic inductances determined from cold measurement at VDS = 0 V, VGS = 0 V, at frequencies from 10 GHz to 20 GHz. (b) Parasitic resistances determined under the same condition. C. Intrinsic Circuit Extraction With the knowledge of extrinsic parasitic parameters, the intrinsic circuit elements shown in Fig 3 can be extracted from the S-parameters measured under the working bias conditions. After de-embedding the extrinsic components, the Y-parameters for the intrinsic part of the device can be expressed as: ⎡ ( jω C gs Ri− 1 ) + ( j ω C dg R dg−1 ) + R −fdg1 + R −fgs1 ⎢ Yint = ⎢ g m e − jωτ − ( jω C dg R dg−1 ) − R −fdg1 ⎢ 1 + j ω C gs Ri ⎣ − ( j ω C dg R dg−1 ) − R −fdg1 R ds−1 + R −fdg1 + j ω C ds + ( j ω C dg ⎤ ⎥ R dg−1 ) ⎥ ⎥ ⎦ (10) with A B = A⋅ B . A+ B The intrinsic differential resistances (Rfgs and Rfdg) are first estimated from Y11 and Y12 at low frequencies (in the megahertz range), since the capacitance terms tend to vanish at such low frequencies. Then, separating the real and imaginary parts in Equation (10) paves the way for 5 determination of the rest of the eight intrinsic parameters [15]. The calculated intrinsic parameters for the same device are listed in Table 2, where ΔS refers to the average error between the calculated and measured S-parameters over the entire frequency range. G Lg Rd Rg Vi Cdg Ids Cgs Ld D Rds Cds Ri Table 2 4.45 Ls (pH) 6.82 Rs (Ω) Ld (pH) 5.47 36.99 Rd (Ω) Lg (pH) 3.69 Rg (Ω) 52.19 69.0 Rfgs (kΩ) 0.91 Cgsp(fF) 62.0 6.84 Cdsp(fF) Rfdg (kΩ) 571.06 Ri (Ω) 8.62 Cgs (fF) 0.00 153.33 Cds (fF) Rds (Ω) 76.44 23.02 Cdg (fF) Rdg (Ω) gm (mS) 41.57 ΔS 9.82% τ (ps) 1.38 The extracted extrinsic parameters from cold-FET measurements, and intrinsic parameters from determined S-parameters measured at VDS = 4 V, VGS = -3 V, frequencies from 2 GHz to 20 GHz. V. HOT-FET EXTRACTION Determination of parasitic circuit elements, as we can see, is imperative for the whole de-embedding procedure. Instead of cold measurements, hot-FET extraction methods attempt to find out the parasitic parameters for the biased conditions. This is necessitated due to the dependence of the FET’s parasitics on bias [18]. Especially for the access resistances, they have been reported to be more critical in GaN HFETs than in other the material systems [9]. Consequently, we examine two hot-FET extraction methods proposed by Manohar et al. and Shirakawa et al. in [10] and [11]. The first method is analytical fitting based while the second is pure optimization based, as elaborated on below. A. Analytical Method For this method, in order to provide the analytical expressions for the circuit elements, the model needs to be simplified as shown in Fig 7. Assuming the pad capacitances are known and de-embedded, the parasitic resistances and inductances can then be correlated with the Z parameters of the rest. Starting from the source pad extraction, it has: Im( Z 12 ) C gd (11) = X 1 + LS ω gm Re( Z 12 ) = C gd gm X 2 + RS with X 1 = sin( ωτ )[Im( Z 21 ) − Im( Z 12 )] + cos( ωτ ) [Re( Z 12 ) − Re( Z 21 )] X 2 = ω cos( ωτ )[Im( Z 21 ) − Im( Z 12 )] − ω sin( ωτ ) [Re( Z 12 ) − Re( Z 21 )] Rs Ls Ids=gmVie-jωτ S S Fig. 7 Circuit model for analytical extraction method proposed in [10] By linearly fitting Im(Z12)/ω on X1 and Re(Z12) on X2, the source Rs and Ls can be extracted by extrapolation. Once Rs and Ls are determined, the rest of the parasitic resistances and inductances can be extracted by polynomial fitting over the frequency on the real and imaginary parts of the following equations: Y Z 12 − Z s (15) = − 12 ,int Z 11 − Z s − Z g Y22 ,int Y Z 12 − Z s = − 12 ,int Z 22 − Z s − Z d Y11 ,int (16) Here Zi = Ri+jωLi (i = g, s, d). Noting that the validity of equations (11) and (12) is based on the assumption that (ωRiCgs)2 ~ 0, the linear regression is best if done for frequencies below 5 GHz. For example, the linear regression in determining the source resistance is illustrated in Fig 8 using the data on the same device. The intrinsic parameters are then calculated after the determination of the parasitics. The transition time τ used for equations (13) and (14) is initially set to be 1ps, and updated iteratively till its convergence satisfied. 30 29.5 Y = 3.58+2.40 X 29 2 2 RS = 3.58 Ohms 28.5 28 (12) 27.5 (13) (14) 27 9.6 9.8 10 10.2 10.4 10.6 10.8 11 X2 Fig. 8 The illustration of analytical method to determine source parasitic resistance. Table 3 shows all the calculated circuit parameters. Here, some of the extracted parasitic resistances and inductances are comparable with the results from cold measurements. However, the parasitic inductances exhibit large variations and fitting errors. The calculated inductances vary significantly when different frequency ranges are used. Sometimes the 6 parasitic values obtained by this method are negative which is non-physical even though extreme care was used in during calibrations and measurements. Besides the possible reason that the pad parasitics may change for hot-FET conditions compared with cold-FET, the omission of the gate differential resistances (Rfdg and Rfgs) , especially when they are in the few kΩ range, may introduce considerable error to the analytical expressions (11) and (12). Re[ Y1 (ω + Δ ω ) − Y1 (ω )] (ω + Δ ω ) Im[ Y1 (ω + Δ ω )] − ω Im[ Y1 (ω )] C gs ⋅ Ri = (18) Moreover we find: R −fgs1 = Re[ Y1 (ω )] − ω Im[ Y1 (ω )] ⋅ (C gs ⋅ Ri ) (19) ω C gs = Im[ Y1 (ω )] ⋅ [1 + (ω ⋅ C gs ⋅ Ri ) ] (20) 2 Table 3 3.58 Ls (pH) 85.54 Rs (Ω) Ld (pH) 6.12 216.72 Rd (Ω) Lg (pH) 0.00 Rg (Ω) 112.65 69.0 Cgsp(fF) 62.0 Cdsp(fF) 415.90 Ri (Ω) 11.07 Cgs (fF) 0.98 215.33 Cds (fF) Rds (Ω) 71.23 Cdg (fF) gm (mS) 28.84 ΔS 14.8% τ (ps) 0.89 The calculated circuit parameters by analytical method, directly from hot-FET data on the same device measured at VDS = 4 V, VGS = -3 V. Pad capacitances obtained from cold measurements. B. Optimization Method In [11], a pure optimization procedure has been developed to extract extrinsic and intrinsic parameters simultaneously for the conventional GaAs HFETs, using the same equivalent circuit shown in Fig 7. After de-embedding the parasitic capacitances, the Z-parameters for the intrinsic and extrinsic part are related by: Z int = Z − Z ext ⎡ R + R s + jω ( L g + Ls ) =Z −⎢ g R s + jω L s ⎣ R s + jω L s ⎤ R d + R s + j ω ( L d + L s ) ⎥⎦ (17) The basic idea of the optimization procedure is to express all of the intrinsic elements as functions of ω and an extrinsic vector, Zext (Rd, Rg, Rs, Ld, Lg, Ls). By assuming that all intrinsic elements are frequency independent, we can obtain the optimized values for both the extrinsic and intrinsic parameters through minimizing their variance over frequency. It should be mentioned that the 7-variable intrinsic circuit is not suitable for a GaN-based device. Especially Rfdg and Rfgs are important in characterizing the current conduction of the gate diode applicable for large signal analysis. We should also note that the traditional optimization procedures can hardly handle the 10-variable intrinsic circuit shown in Fig 3. This has its genesis in the fact that it is impossible to express all 10 variables solely as a function of ω in light of the fact that there are only 4 measured pieces of data, including the real and imaginary parts, at each frequency. Our approach is to find expressions for each of the intrinsic circuit elements in terms of the intrinsic Y-parameters at multiple frequencies. In other words, we chose a proper ∆ω so that at each frequency ω, all the elements are determined from both Y(ω) and Y(ω+∆ω). Taking the Rfgs, Ri, Cgs sub-circuit as an example: where Y1 = Y11,int+Y12,int. Similarly, the Rfdg, Rdg, Cdg sub-circuit follows: C dg ⋅ R dg = Re[ Y2 (ω + Δ ω ) − Y2 (ω )] (ω + Δ ω ) Im[ Y2 (ω + Δ ω )] − ω Im[ Y2 (ω )] (21) R −fdg1 = Re[ Y2 (ω )] − ω Im[ Y2 (ω )] ⋅ ( C dg ⋅ R dg ) (22) ω C dg = Im[ Y2 (ω )] ⋅ [1 + (ω ⋅ C dg ⋅ R dg ) 2 ] (23) where Y2 = -Y12,int. When programming, we also smoothed the calculated values by averaging them within several adjacent data points in order to reduce the scatter caused by any error in the measurement. The rest of the 4 intrinsic parameters are expressed as follows: R ds− 1 = Re( Y22 ,int + Y12 ,int ) (24) ω C ds = Im( Y22 ,int + Y12 ,int ) (25) g m = (Y21 ,int − Y12 ,int ) ⋅ [1 + j ω (C gs ⋅ Ri )] (26) Im{( Y21 ,int − Y12 ,int ) ⋅ [1 + j ω ( C gs ⋅ Ri )]} ωτ = tan −1 Re{( Y21 ,int − Y12 ,int ) ⋅ [1 + j ω (C gs ⋅ Ri )]} (27) The variance of each intrinsic element is: N εi = ∑ [ f (ω k =1 i k , Z ext ) − f i (ω , Z ext ) ] 2 N −1 ( i = 1, 2 , " 10 ) (28) Here fi represents the analytical expressions of each intrinsic parameter obtained from Equations (18) through (27). We define a global weighted scalar as the objective function for the optimization routine: 10 ε = ∑ Wiε i (29) i =1 The weighting factors are selected from the normalization of minimum variance according to the optimization procedure performed on each intrinsic parameter. In another word, if ℜ(⋅) denotes the optimization routine, its output will give both the minimized variance and corresponding optimized 7 extrinsic vector: ℜ( fi , ε i , Z Table 4 initial ext ) → [εˆi , Z optimized ext (30) ] Then Wi is defined as: Wi = 1 (31) εˆi To avoid any non-physical convergence, we also strengthen the optimization routine by imposing the following inequality constraints: 2.80 Ls (pH) 81.55 Rs (Ω) Ld (pH) 3.94 180.91 Rd (Ω) Lg (pH) 0.00 Rg (Ω) 13.61 69.0 Rfgs (kΩ) 0.93 Cgsp(fF) 62.0 22.96 Cdsp(fF) Rfdg (kΩ) 489.26 Ri (Ω) 3.30 Cgs (fF) 0.00 197.17 Cds (fF) Rds (Ω) 76.56 72.72 Cdg (fF) Rdg (Ω) gm (mS) 32.92 ΔS 3.86% τ (ps) 0.51 The calculated circuit parameters by optimization method, directly from hot-FET data on the same device measured at VDS = 4 V, VGS = -3 V. Pad capacitances obtained from cold measurements. 200 Rdg without Rfdg [ R d , R g , " , L s ] ≥ 0 and [ f1 , f 2 , " , f10 ] ≥ 0 (32) Rdg = -Re(1/Y 12, int ) 150 Rdg with Rfdg A sequential quadratic programming (SQP) algorithm [19], which is a generalization of Newton’s method, is utilized in solving this constrained optimization problem. The objective function and the constraints are replaced with the quadratic and linear approximation, respectively. The convergence properties of the algorithm can be improved by using a line search with penalty parameters recommended in [20]. Noting that it is a gradient-based method, the convergence of the program relies on the continuity of the objective function. Therefore, the measurements should be carefully conducted to avoid large noise or the moving of the reference planes. (a) 100 R without R i 50 15 fgs Ri with Rfgs 0 0 0 5 10 15 20 Frequency (GHz) 600 Cgs 400 Cold pinch-off measurement Hot measurement (b) Start Optimization 200 De-embedding capacitances contribution Extract Cgsp, Cdsp C 0 Find intrinsic element fi 10 ∑W ε i =1 i i as objective function Run optimization Obtain optimized Zext Calculate intrinsic parameters Evaluate S-parameters Update Zext dg Cds Calculate intrinsic Zand Y-parameters For i = 1~10 Set εi as objective function Run optimization for fi Obtain Wi Set ε = Set initial value for Zext and ∆ω -200 0 5 10 15 Frequency (GHz) 20 Calculate objective function 40 Objective < Criterion 15 N Y Output data End 10 30 (c) 5 Fig. 8 Flow chart diagram for the optimization extraction procedure. In short, our optimization routine first finds the weighting factors by minimizing the variance for each intrinsic element over frequency. Then the weighted average error is set to be the objective function, and the routine is run again to obtain the globally optimized extrinsic vector. Finally, all intrinsic parameters are calculated and the S-parameters are evaluated. The flow chart demonstrating this extraction procedure is shown in Fig 8. The calculated intrinsic and extrinsic parameters are listed in Table 4. 20 0 10 -5 0 5 10 15 Frequency (GHz) 20 8 1/3S 21 5S 12 S 11 S22 (d) Fig. 9 Optimized intrinsic elements from the data measured at VDS = 4 V, VGS = -3 V, and frequencies from 2 GHz to 20 GHz. (a) Ri and Rdg, with and without the differential resistances Rfdg and Rfgs considered. (b) Cdg, Cgs, and Cds. (c) gm and τ. (d) Measured (×) and simulated (line) S-parameters In Fig 9, we display some of the intrinsic parameters determined by the routine. The calculated intrinsic parameters have shown good frequency invariance properties. In order to demonstrate the importance of the differential resistance values, we also optimized and simulated a simplified intrinsic circuit model without the differential resistances. Neglecting them imparts a profound effect onto Rdg and Ri at lower frequencies as shown in Fig 4 (a). A good agreement is achieved between measured and simulated S-parameters with only a 3.86% average error as indicated in Table 4, in 2 ~ 20 GHz frequency range with VDS = 4 V, VGS = -3 V. conditions. Also for GaN HFETs, the effect of material defects be also involved since the frequency range used for extrinsic element extraction between hot and cold methods are different. The error generated in the cold measurements could propagate, affecting the accuracy of the intrinsic parameters obtained thereafter. From a mathematical point of view, the optimization-based extraction routine will undoubtedly produce the smallest error at one given bias condition. The method is more practicable if its stability can be verified by testing whether the results depend on the initial input vector. By selecting all zeroes, random numbers, and results from the cold measurement as the initial values for Zext, we found that for each bias, the routine led to the same results regardless of the initial values chosen. Furthermore, it is more meaningful to evaluate the validity of this method at various bias conditions. Thus, 15 different bias conditions were tested for one device as shown in Table 5. The overall simulation error is seen as approximately 5%. In contrast, this error increases to more than 10% for the corresponding bias points using the constant extrinsic parameters from cold measurements. In large signal device nonlinearity analysis, it is usually convenient to assume the extrinsic parameters to be constant. If one were to do so by taking the mean value of the extrinsic parameters over all bias points as the set of bias-independent extrinsic parameters, the error will increase, remaining below 9% for every bias condition. Table 5 VGS (V) VDS (V) 3.0 VI. DISCUSSIONS AND CONCLUSIONS Arguably, the equivalent circuit for a GaN HFET is slightly more complicated than the conventional models for GaAs or SiC devices. Consideration of the gate differential resistances is necessary, which under the reverse bias conditions represent leakage current paths of the gate Schottky diode. Due to the defects and dislocations in the crystal and assistance from surface states or traps, the leakage current term is always non-negligible. Especially at low frequencies where the value of ωC is small, the differential resistance term will dominate the total conductance. Although the increase in the number of circuit elements brings extra complexity in hot-FET extraction, the intrinsic parameters can still be calculated via the strengthened optimization routine. By analyzing and comparing different small signal modeling methods, we could examine the validity and consistency of the results. In our GaN HFET device, the parasitic resistances extracted under different techniques are consistent, while the parasitic inductances show large variation. Basically, the inductances calculated from hot-FET models are always larger than that from cold-FET measurements. This difference may reflect the change of the pad parasitics under hot bias conditions compared with cold 4.0 5.0 6.0 7.0 -2.5 -3.0 -3.5 8.21, 0.93, 3.32 4.47% 2.72, 11.67, 3.95 5.47% 4.36, 6.32, 2.19 4.32% 0.00, 14.67, 3.63 5.42% 0.00, 4.19, 0.00 2.64% Rs , Rd , Rg (Ω) ΔS 8.34, 3.26, 4.19 6.71% 2.80, 3.94, 0.00 3.86% 0.00, 4.45, 3.58 5.32% 2.14, 6.65, 1.05 5.46% 0.00, 11.24, 2.44 5.76% Rs , Rd , Rg (Ω) ΔS 0.00, 8.73, 2.64 4.00% 3.43, 7.78, 1.13 5.73% 0.00, 4.83, 3.58 5.39% 2.20, 9.09, 2.50 5.34% 0.00, 0.00, 2.89 4.50% Rs , Rd , Rg (Ω) ΔS Average: Rs = 2.28 Ω, Rd = 6.52 Ω, , Rg =2.47Ω The parasitic resistances and calculated S-parameters simulation error for different bias conditions, from optimization method. In all, we proposed an 18-element equivalent circuit model to determine the small signal parameters for AlGaN/GaN HEFTs. We find that the differential gate resistances must be included in the intrinsic part of the circuit to take into account the current conduction through the gate diode. Different extraction methods were examined including the conventional cold-FET measurements and hot-FET varieties. A hybrid extraction method, which combines the cold-FET measurements to obtain parasitic capacitances and hot-FET 9 data to calculate parasitic resistances/inductances, is developed and capable to give the highest accuracy with an overall error of approximately 5% without any dependence on the initial trial value. The model utilizes data taken at multiple frequencies to extract the intrinsic parameters. Additionally, non-negative values for all the elements are imposed for the nonlinear optimization routine. The program searches the optimized extrinsic resistances/inductances by minimizing the weighted variances of the intrinsic parameters, with the weighting factors determined from a normalization process. The extrinsic inductances show a notable difference between the different methods. We believe that the inductances extracted from the conventional cold measurements are not accurate enough under operating bias conditions. The values obtained from the conventional cold measurements should be thoroughly validated by other extraction methods. Our hybrid optimization routine can generate more reliable extrinsic and intrinsic values for the small signal equivalent circuit for AlGaN/GaN HFETs. [13] [14] [15] [16] [17] [18] [19] [20] [21] ACKNOWLEDGMENT This work has been funded by a grant from the Air Force Office of Scientific Research under the direction of Drs. K. Reinhardt and D. J. Silversmith. REFERENCES [1] H. Morkoç, “Handbook of Nitride Semiconductors and Devices”, Volume I, II, III Wiley-VCH, 2008, ch. 1-3. [2] B. Hughes and P. J. Tasker, “Bias dependence of the MODFET intrinsic model elements values at microwave frequencies,” IEEE Trans. Electron Devices, vol. 36, No. 10, pp. 2267–2273, Oct. 1989. [3] G. Dambrine, A. Cappy, F. Heliodore, and E. Playez, “A new method for determining the FET small-signal equivalent circuit”, IEEE Trans. Microw. Theory Tech., vol. 36, No. 7, pp. 1151-1159, Jul. 1988. [4] G. Chen, V. Kumar, R. S. Schwindt, and I. Adesida,“A Low Gate Bias Model Extraction Technique for AlGaN/GaN HEMTs”, IEEE Trans. Microw. Theory Tech., vol.54. No. 7, pp.2949-2953, Jul. 2006. [5] R. G. Brady, C. H. Oxley, and T. J. Brazil, “A improved small-signal parameter-extraction algorithm for GaN HEMT devices”, IEEE Trans. Microw. Theory Tech., vol.56. No. 7, pp.1535-1544, Jul. 2008. [6] A. Jarndal and G. Kompa, “A new small-signal modeling approach applied to GaN devices”, IEEE Trans. Microw. Theory Tech., vol. 53, No. 11, pp. 3440-3448, Nov. 2005. [7] G. Meneghesso, G. Verzellesi, R. Pierobon, F. Rampazzo, A. Chini, U. K. Mishra, C. Canali, and E. Zanoni, “Surface-related drain current dispersion effects in AlGaNGaN HEMTs,” IEEE Trans. Electron Devices, vol. 51, No. 10, pp. 1554-1561, Oct. 2004. [8] A. E. Parker and J. G. Rathmell, “Broad-band characterization of FET self-heating,” IEEE Trans. Microw. Theory Tech., vol. 53, No. 7, pp. 2424-2429, Jul. 2005. [9] David W. DiSanto and C. R. Bolognesi, “At-Bias Extraction of Access Parasitic Resistances in AlGaN/GaN HEMTs: Impact on Device Linearity and Channel Electron Velocity”, IEEE Trans. on Electron Devices, vol. 53, No. 12, pp. 2914-2919, Dec. 2006. [10] J. Burm, W. Schaff, L. Eastman, H. Amano, and I. Akasaki, “An improved small signal equivalent circuit model for III-V nitride MODFET’s with large contact resistances,” IEEE Trans. on Electron Devices, vol. 44, No. 5, pp. 906-907, May 1997. [11] S. Manohar, A. Pham, and N. Evers, “Direct determination of the bias-dependent series parasitic elements in SiC MESFETs”, IEEE Trans. Microw. Theory Tech., vol. 51, No. 2, pp. 597-600, Feb. 2003. [12] K. Shirakawa, H. Oikawa, T. Shimura, T. Kawasaki, Y. Ohashi, T. Saito, and Y. Daido, “An approach to determining an equivalent circuit for HEMTs”, IEEE Trans. Microw. Theory Tech., vol. 43, No.3, pp. 499-503, Mar. 1995. C. Tsironis and R. Meierer, “Microwave wide-band model of GaAs dual gate MESFETs”, IEEE Trans. Microw. Theory Tech., vol. 30, No. 3, pp. 243-251, Mar. 1982. R. G. Brady, C. H. Oxley, and T. J. Brazil, “An Improved Small-Signal Parameter-Extraction Algorithm for GaN HEMT Devices” IEEE Trans. Microw. Theory Tech., vol. 56, No. 7, pp. 1535-1544, Jul. 2008. M. Berroth and R. Bosch, “High-frequency equivalent circuit of GaAs FETs for large-signal applications”, IEEE Trans. Microw. Theory Tech., vol.39. No. 2, pp.224-229, Feb 1991. E. Chigaeva and W. Walthes, “Determination of small-signal parameters of GaN-based HEMTs,” in Proc. IEEE/Cornell High Performance Devices Conf., 2000, pp. 115–122. M. Berroth and R. Bosch, “Broad-Band Determination of the FET Small-Signal Equivalent Circuit”, IEEE Trans. Microw. Theory Tech., vol.38. No. 7, pp.891-895, Feb 1990. C. F. Campbell and S. A. Brown, “An analytic method to determine GaAs FET parasitic inductances and drain resistance under active bias conditions,” IEEE Trans. Microw. Theory Tech., vol. 49, no. 7, pp. 1241– 1247, Jul. 2001. The Mathworks Inc, “Online Documentation for Mathworks Products, Optimization Toolbox” M. J. D. Powell, “A fast algorithm for nonlinearly constrained optimization calculations”, Springer Berlin, Volume 630, 1978. Y. A. Khalaf, S. M. Riad, “Novel Technique for Estimating Metal Semiconductor Field Effect Transistor Parasitics”, International Journal of RF and Microwave Computer Aided Engineering, Vol. 13, 1, pp. 62-73, Dec. 2002. Qian Fan received his B.S and M.S degrees from Information and Electronic Engineering Department, Zhejiang University, Hangzhou, China. In 2004, he entered the Electrical and Computer Engineering Department in Virginia Commonwealth University and now is working toward the Ph.D degree. His current research interests include GaN-based device fabrication; structure growth (by MOCVD); device characterizations and modeling. Jacob H. Leach (Student Member, IEEE) received B.S degrees in physics and electrical engineering with honors from Virginia Commonwealth University in 2004. He earned the M.S. degree, also from Virginia Commonwealth University in electrical engineering in 2007. His current research interests are in high frequency and transient responses of GaN-based devices and their applications. Hadis Morkoç received the B.S.E.E. and M.S.E.E. degrees from Istanbul Technical University, Istanbul, Turkey, and the Ph.D. degree in electrical engineering from Cornell University, Ithaca, NY. From 1976 to 1978, he was with Varian Associates, Palo Alto, CA, where he was involved in various novel FET structures and optical emitters based on then new semiconductor heterostructures. He held visiting positions at the AT&T Bell Laboratories (1978–1979), the California Institute of Technology, Pasadena, and Jet Propulsion Laboratory (1987–1988), and the Air Force Research Laboratories-Wright Patterson AFB as a University Resident Research Professor (1995–1997). From 1978 to 1997, he was with the University of Illinois, Urbana. In 1997, he joined the newly established School of Engineering, Virginia Commonwealth University, Richmond. He and his group members have been responsible for a number of advancements in 10 compound semiconductor, including wide-bandgap nitride, heterostructures, and devices. He has been a prolific writer with a number of books, review and tutorial papers, book chapters, and journal publications. He is among the most cited in the fields of engineering, physics, and materials science. Dr. Morkoç is a Fellow of the American Association for the Advancement of Science, a Life Fellow of the American Physical Society, and was a Fellow of the IEEE through 2008, a member of Sigma Pi Sigma and Eta Kappa Nu, and a Life Member of Sigma Xi and Phi Kappa Phi. He is listed in Who’s Who in America, Who’s Who in the Midwest, American Men and Women in Science, Who’s Who in Engineering, and International Men of Achievement.