Largesignal mmwave InAlN/GaN HEMT Power Amplifier Characterization through Selfconsistent Harmonic Balance / Cellular Monte Carlo Device Simulation D. Guerra, F. A. Marino†, D. K. Ferry, S. M. Goodnick, M. Saraniti , and R. Soligo Center for Computational Nanoscience, Arizona State University, 551 E Tyler Mall, Tempe, AZ, 85287, USA † also with: Department of Information Engineering, via Gradenigo 6/B, 35131, Padova, Italy Tel: +1 480 9652650, Fax: +1 480 9653837, Email:
[email protected] Abstract
Harmonic Balance / Cellular Monte Carlo Selfconsistently coupled CircuitDevice Simulations
We report the simulation of the largesignal performance of mmwave FET power amplifiers obtained for the first time through Full Band Monte Carlo particlebased device simulation selfconsistently coupled with a Harmonic Balance (HB) frequency domain circuit solver. Due to the iterative nature of the HB algorithm, this FET simulation approach is possible only due to the computational efficiency of our Cellular Monte Carlo (CMC), which uses precomputed scattering tables. On the other hand, a frequency domain circuit solver such as HB allows the simulation of the steadystate behavior of an external passive reactive network without the need for simulating long transient time (i.e. RC, L/C time constants) typical of time domain solutions. By exploiting this newly developed selfconsistent CMC/HB code, we were able to timeefficiently characterize the mmwave power performance of a stateoftheart 30nm gatelength InAlN/GaN HEMT. Introduction Smallsignal AC analysis fails to analytically predict largesignal device performance, which must be assessed by simulating the device within the full range of actual operating conditions. Thus, the power amplifier highQ (i.e. highly selective in frequency) matching network must be included in order to properly simulate the dynamic loadline seen at the device output. A time domain solution of these matching networks including large reactive elements implies long transient times. However, this issue can be overcome by connecting a sinusoidal voltage generator, tuned at the fundamental frequency, at the device output [1] (i.e. an active loadline). This emulates the loadline drain voltage swing at the fundamental harmonic, and presents a shortcircuit at the other harmonics, effectively emulating a highQ matching network as shown in Fig.1. The actual synthesized load impedance is determined in postprocessing through Fourier transform. The magnitude and phase of the complex load can be also adjusted, by changing magnitude and phase of the voltage generator, and the simulation and the subsequent impedance analysis can be iterated until the desired load impedance is obtained. In such way, we can emulate a constant load for different input powers, and characterize a device under largesignal operations as shown in Fig. 2, 3, and 4.
9781457705052/11/$26.00 ©2011 IEEE
The iterative procedure described in the previous paragraph is a particular oneharmonic case of the general frequencydomain circuit solver known as Harmonic Balance (HB) [2]. An automated version of the described procedure has been implemented, selfconsistently coupled with our Cellular Monte Carlo (CMC) [3] device simulator, and extended to a variable number of harmonics allowing us to emulate any kind of impedance and network connected to any of the device contact as shown in the example in Fig.5, 6, and 7. In this figure, a Finite Difference Time Domain (FDTD) solver was also developed and selfconsistently coupled to the CMC in order to compare this timedomain circuit solution with the frequencydomain circuit solution of HB. Unlike previously MC codes only coupled with timedomain circuit solvers used for FET power analysis [4], our CMC/HB simulator allows a timeefficient simulation of devices connected to highQ matching networks (required to suppress undesired harmonics generated due to nonlinearity) typical of mmwave band power amplifier classes such as AB, B, F, and harddriven ClassA. In general, the frequency domain circuit solution provided by HB allows including the effect of those passive structures that are more easily characterized in the frequency domain than in the time domain. For instance, the HB algorithm can include the effect of a transmission line by using the Sparameters characterization, which still holds even under largesignal operations for a passive and bias independent structure with a linear response. InAlN/GaN HEMT Power Amplifier Characterization Using this approach, we characterized the RF power performance of a stateoftheart InAlN/GaN HEMT recently reported by Lee et al [5]. We first performed a fit of the experimental DC characteristics as show in Fig.8. Then, the RF smallsignal performance were obtained through CMC/AC simulations (by applying small step perturbations to extract the Yparameters), and the agreement with the experimental fT is shown in Fig.9. This device has a relatively low experimental fMAX value of 13 GHz due to the high gate resistance, RG, of the rectangular gate. On the other hand, Lee et al. also reported that a mushroomshaped gate structures or multifinger devices with rectangular gates can significantly improve this fMAX value. The effect of RG cannot be directly taken into account
34.2.1
IEDM11793
by the solely Monte Carlo device simulator. However, RG, modeled as a lumped series resistor, can be included in our characterization by using different techniques. In particular, smallsignal analysis allows to analytically embedding in postprocessing the impact of RG in the smallsignal twoport network characterization of the device through the Yparameters. The impact of RG on fMAX can be seen in Fig. 9. On other hand, the HB and/or the FDTD solver, selfconsistently coupled with the CMC, allows including the effect of RG under largesignal operations, where an analytical parameterization is not available, through selfconsistently coupled circuitdevice simulations. Thus, in our simulated characterization we used different values of gate resistance, simulating in such way what the performance of this device would be if a low resistance gate structure were used. In general, parasitic elements can be simulated in the timedomain due to the fact that the reactive values are in general small, resulting in small transient time. Therefore, an iterative circuit solver such as HB is not essential in this case. The largesignal RF characterization was performed in the Kaband with a 35 GHz singletone continuouswave with the device biased for ClassA operations, and a 50 load was emulated as example. The actual load that maximizes the output power can be obtained from experimental loadpull measurements as well as by computational loadpull that can be readily obtained through the active loadline technique [1]. Finally, we proceeded with the CMC/HB simulations including some real device issues such as parasitic elements (gate resistance, Fig.10), and material defects/reliability such as threading dislocations (Fig.11) and surface traps (Fig.12). The gate resistance was included by combining the HB and the FDTD circuit solvers. HB timeefficiently took into account the effect of the highQ output matching network and loadline, while FDTD simulated the impact of RG. The surface traps were modeled by using a sheet of negative charge depleting the 2DEG. This approach cannot take into account the timevarying trapping/detrapping process, but corresponds to the worstcase where all the traps are filled with electrons.
References [1]. O. Bengtsson, L. Vestling, and J. Olsson, “A computational loadpull method with harmonic loading for highefficiency investigations,” SolidState Electronics, vol. 53, no. 1, pp. 86 – 94, 2009. [2]. S. A. Maas, Nonlinear Microwave and RF Circuits, 2nd ed. Norwood, MA: Artech House, 2003. [3]. M. Saraniti and S. Goodnick, “Hybrid fullband Cellular Automaton/Monte Carlo approach for fast simulation of charge transport in semiconductors,” IEEE Transactions on Electron Devices, vol. 47, no. 10, pp. 1909–1915, October 2000. [4]. H. I. Fujishiro, S. Narita, and Y. Tomita, “Large signal analysis of AlGaN/GaNHEMT amplifier by coupled physical devicecircuit simulation,” Physica Status Solidi (a), vol. 203, no. 7, pp. 1866 – 1871, 2006. [5]. D. S. Lee, J. Chung, H. Wang, X. Gao, S. Guo, P. Fay, and T. Palacios, “245GHz InAlN/GaN HEMTs with oxygen plasma treatment,” IEEE Electron Device Letters, vol. 32, no. 6, pp. 755–757, June 2011.
Conclusions
VDD
ClassB
vD = VDD + vd vTH
id: f1, f2, f4, ...
freq
vL: f1 ZL’
= vd: f1
vG: f1 ZL’
VDD
+ _
Figure 1: Example of a ClassB amplifier (i.e. conducting only for half of the input cycle) with highQ matching network emulated by an active loadline technique. f = 25 GHz ClassB
drain current [mA/mm]
ZL'= 50 Ω
3500
POUT=6.1 W/mm
3000
POUT=5.6 W/mm POUT=4.7 W/mm
2500
POUT=3.2 W/mm
2000
POUT=1.7 W/mm
1500 1000
POUT=0.6 W/mm
500 0
Acknowledgment
0
Work supported as part of the Air Force Research Laboratory (AFRL) National High Reliability Electronics Virtual Center (HiREV) under Wyle Laboratory Contract #DD8192 (Monitor: S. Tetlak).
IEDM11794
ZL
HighQ Matching Network
DC BIAS: Vg= 6 V, Vd=18 V
In summary, we demonstrated that the selfconsistent combination of a frequency domain circuit solver such as Harmonic Balance with the Full Band Cellular Monte Carlo device simulator allowed us to successfully perform for the first time the largesignal RF characterization of highQ matched power amplifiers by using a particlebased device simulator. The same set of simulations with timedomain and/or conventional Monte Carlo (MC) (Fig.13) techniques would have required a prohibitive simulation time.
id: f1
vG: f1
vD = VDD + vd
vTH
f1 f2 f3 f4 f5
S11
34.2.2
10
20
30
40
time [ps]
50
60
70
Figure 2: Resulting simulated drain current waveforms obtained for a ClassB power amplifier by performing an input power sweep with constant emulated load impedance.
2500
f = 25 GHz ClassA
DC BIAS: Vg= 2 V, Vd=18 V
Pout [dBm], Gain [dB]
2000
1500
1000
500
0
5
10
15
20
drain voltage [V]
25
vG
i : f , f , f , ... = v : f , f , f , ...
GND
S
1
2
S
1
2
20 15
GND
iS: f1, f2, f3, ... vS: f1, f2, f3, ... f1
3
f2
LS
10
5
5 5
10
15
20
0
25
CMC current HB current
ClassB
0.05
vG
3
15
Gain
10
Pin [dBm] Figure 4: Typical largesignal figures of merit of the power performance of the test device of Fig.3 obtained through CMC/HB. vD 0.1
iD f1
25
20
0
Figure 3: Example of the CMC/HB simulated dynamic loadlines of an AlGaN/GaN HEMT test device biased as ClassA power amplifier with increasing input power. vD
f1
30
PAE
P1dB
25
30
iD
out
1dB
source current [A]
0
30
0 0
source voltage [V]
drain current [mA/mm]
35
[6].
IdVd: Vg = 6.0 : 0.5 : 1.0 V WG = 100 um
f = 25 GHz ClassA 40 WG = 100 um 35 P
ZL'= 100  j20 Ω
PAE [%]
DC BIAS: Vg= 2 V, Vd=18 V ZL'= 100  j20 Ω
20
0.1
60
80
60
80
FDTD/CMC voltage HB/CMC voltage
0.05
f3
40
time [ps]
0
0.05 0
20
40
time [ps]
HB: 3 harmonics HB current HB harmonics
ClassB
0
0.1
20
40
time [ps]
60
0.05 0 0.05 20
40
time [ps]
60
Figure 6: Sinusoidal waveforms of the first three harmonics generated by the Harmonic Balance algorithm.
ClassB
0.05 0.04
0.03
0.03
0.02
0.02
0.01
0.01 0
50
100 frequency [GHz]
150
0
0.05
0.05
0.04
0.04
0.03
0.03
0.02
0.02
0.01
0.01
0
80
ZS (0, 50, 100, 150 GHz ) = 0, 1, 2, 3 Ω
0.04
0
80
HB voltage HB harmonics
0
f = 50 GHz, LS = 1 pH
0.05
0.05
0
source voltage [V]
f = 50 GHz, LS = 1 pH
current [A]
0.1
voltage [V]
source current [A]
Figure 5: Example of the general HB algorithm: the source inductor is emulated by generating the voltage sinusoids at each harmonic. A FiniteDifference TimeDomain (FDTD) circuit solver also selfconsistently coupled with the CMC is used for comparison. The FDTD solver uses a moving average filter that reduces the noise but adds a phase delay in the solution.
0
50
100 frequency [GHz]
150
0
Figure 7: Harmonic content of the analyzed source current, and of the voltage waveform generated by the Harmonic Balance algorithm. 34.2.3
IEDM11795
SOURCE
635 nm 30 nm
2000
635 nm
GATE
Al2O3 In0.17Al0.83N AlN
10 nm 4.5 nm 1 nm
20
CMC simulations Experimental
DRAIN
2500
fmax= 25, 52, 97, 143, 192 GHz
UPG
GaN
300 nm
SELFHEATING
Gain [dB]
1500
1000
fT=268GHz
10
5
500
Increasing RG:
0
1
2
3
4
5
6
7
8
9
40, 70, 150, 500, 2K Ω 0 0 1 10 10
10
drain voltage [ V ]
Figure 8: Layout of the InAlN/GaN HEMT simulation domain and agreement between the DC simulated (without thermal correction) and experimental results [5]. DC BIAS: Vg= 2 V, Vd=15 V WG = 100 um ZL'= 50 Ω
f = 35 GHz ClassA 30
RG=0, 40, 70 Ω
20
15
15 increasing RG
Gain
10
10
5 0
Pout [dBm], Gain [dB]
increasing RG
20
30
25
PAE
increasing RG
5 5
10
15
20
20 no, GD, SG&GD
15
no, GD, SG&GD
10
Gain
5 0
20 increasing NDIS
15
10
15
20
10
10 5
5
10
15
20
25
150

160 ps of simulated time 4 sine periods at 25 GHz = one HB iteration 50000 15 particles 2.5x1016s Poisson step 2.0x10 s Free Flight
100
50
CMC 0
25
0
100
200
300
400
500
Field Plate Length [nm]
Pin [dBm] Figure 12: Effects of surface traps on the ClassA RF power amplifier performance of the device. The traps were located in the gatedrain region (GD) and in both the sourcegate and gatedrain regions (SG&GD) IEDM11796
Gain
5
5 5
15
increasing NDIS
10
25
CPU time [hours]
20
20
MC PAE [%]
Pout [dBm], Gain [dB]
PAE
no, GD, SG&GD
25
PAE
increasing NDIS
200
Pout
25
Pout
Figure 11: Effects of threading dislocations on the ClassA RF power amplifier performance of the device.
f = 35 GHz ClassA 30
NT=3x1012 cm2
15
NDIS=1, 50, 150,300,500 x109 cm
Pin [dBm]
Figure 10: Effects of the gate resistance on the ClassA RF power amplifier performance of the device. DC BIAS: Vg= 2 V, Vd=15 V WG = 100 um Z L'= 50 Ω
f = 35 GHz ClassA 30 3
25
0
25
Pin [dBm]
30
DC BIAS: Vg= 2 V, Vd=15 V WG = 100 um Z L'= 50 Ω
Pout
25
2
Figure 9: AC smallsignal Monte Carlo simulations. The experimental fT is 245 GHz [5]. RG was analytically embedded in the Yparameters by postprocessing.
PAE [%]
Pout [dBm], Gain [dB]
30
10
Frequency [GHz]
PAE [%]
0
h21
15
2DEG
Figure 13: MC and CMC simulation time (1HB iteration) vs increasing field plate length (i.e. decreasing peak electric field, peak scattering, and number of scattering events to mange). HB average number of iterations is 4. 34.2.4