A wideband RF power amplifier (PA) is an essential component in many RF applications, such as electronic warfare (EW) systems and security communications. In the past, the traveling wave tube amplifier (TWTA) has been most commonly used to achieve wideband output power above watt levels [1,2]. However, the emergence of GaN PAs raises the prospect of replacing bulky TWTAs with compact solid state power amplifiers (SSPAs) [3]. GaN high-electron-mobility transistors (HEMTs) have several inherent advantages, including high breakdown voltage, high current density, and high saturation velocity resulting from their wide band gap properties [4]. Unfortunately, unlike GaAs or InP HEMTs, GaN HEMTs have a large power dissipation, which causes a prominent self-heating phenomenon that degrades the RF performance of devices [5]. Thus, the self-heating effect must be considered when designing wideband RF GaN PAs. An accurate large-signal model that includes a thermal model therefore becomes an inevitable requirement. Large-signal models provided by the foundry service are still not mature and do not guarantee model accuracy up to frequencies close to the maximum oscillation frequency (F_max). In-house large-signal modeling should definitely be performed, based on measured data.

The optimum output load point of an RF PA typically shows severe variation with the frequency. Because of this variation, a conventional output load matching prevents RF PAs from obtaining wideband output power. The output load matching circuits should therefore be designed by taking into consideration the frequency-dependent load variation.

This paper describes the development of in-house large-signal models of a GaN HEMT that include a thermal model. Distributed L-C output load matching is proposed for the design of reactive matched amplifiers. Section II presents the large-signal models, followed by the proposed design of the reactive matching PA in Section III. Experimental results are presented in Section IV.

Our in-house models are based on an Angelov model [6,7]. Fig. 1 shows an Angelov model-based GaN HEMT large-signal equivalent circuit including a thermal model. Our inhouse models consist of the Angelov model library supported by Keysight’s ADS 2013 program and thermal sub-circuits [8]. Thermal sub-circuits inform a temperature-dependent large-signal model of the channel temperature raised by R_th. In particular, our in-house models incorporate a nonlinear drain current (I_ds) that is divided into a nonlinear drain DC current (DC I_ds) and a nonlinear drain RF current (RF I_ds). This expression can make the I_ds models reflect the frequency dispersion effect [9]. At high frequencies above a few megahertz, the RF I_ds is activated through virtual inductances in the equivalent circuit model. Equations for the DC I_ds are defined in (1)–(10). Each model parameter is optimized and fitted in comparison with measured DC–IV curves and S-parameters by ADS 2013. The RF I_ds uses the same equation as the DC I_ds. Nonlinear capacitances (C_gs, C_gd, and C_ds) are fitted to well-known Angelov’s nonlinear capacitance model equations [6].

Fig. 2 shows the large-signal modeling procedures. First, DC–IV curves are measured in a device under test (DUT) by an HP 4142B DC source and a Keysight’s IC-CAP program. The sample dies of GaN HEMT, provided by the foundry service, are used as DUTs. The S-parameter is then measured by a vector network analyzer (VNA). The extrinsic parameters and intrinsic parameters are extracted separately by performing hot and cold measurements, respectively. The small-signal model parameters are then extracted from the measured S-parameters according to multiple biases. Nonlinear large-signal model parameters, such as gate-source capacitances (C_gs), gate-drain capacitances (C_gd), drain-source capacitance (C_ds), DC I_ds, and RF I_ds, are modeled by Angelov model-based tangent-hyperbolic equations. In addition, the thermal effect is reflected by performing a pulsed IV measurement using a DIVA D265 instrument and a thermal chuck. A thermal resistance (R_th) is extracted using the pulsed IV measurement method [8]. The DC I_ds is fitted, including the extracted R_th, to represent negative current slopes by the thermal effect under high drain voltage and high drain current regions. The RF I_ds is measured under a quiescent bias condition and fitted to the RF I_ds equations. An initially-equipped large-signal model is verified by comparison with the measured DC I_ds, RF I_ds, and S-parameters. Finally, the complete model is optimized by some iterations of the above procedures. The size of the GaN HEMT for large-signal modeling is selected as 6 (fingers) × 125 μm (gate width) under the criteria of F_max over 20 GHz and load pull power over 34 dBm from 6 to 18 GHz. The extracted I_ds model parameters are summarized in Table 1.

Fig. 3 shows a comparison of the DC–IV curves between the measurements and the models, measured at drain voltages (V_ds) of 0 to 36 V and gate-source voltages (V_gs) of -3.6 to -1.6 V. The in-house model shows better agreement with the measured data than is observed with the model provided by the foundry service. Fig. 4 shows an RF I_ds according to temperature. Under 24 V of a quiescent drain voltage (V_dsq) and -2.0 V of a quiescent gate voltage (V_gsq), a pulsed I_ds is measured and an RF I_ds equation is fitted. The pulse width is 200 ns. The in-house model predicts a reduced RF I_ds when temperature increases.

Fig. 5 represents the comparison between the measured S-parameters and the simulated S-parameters from 0.5 to 60 GHz. As shown in Fig. 5, the in-house model shows better agreement with the measured data up to high frequencies at S₂₂ when compared to the model provided by the foundry service. This results from the exact RF I_ds modeling by the pulsed IV measurement. Finally, a load pull data at 10 GHz is compared between the measurement and the model. As shown in Fig. 6, the in-house model predicts a more realistic output power than is obtained with the model provided by the foundry service. This result confirms the validity of our large-signal model that includes a thermal model.

A reactive matched PA is a type of wideband power amplifier that has an octave bandwidth. It purposely decreases the Q-Fig factors by inserting lossy matching components at the gate of the transistors [11-14]. This approach improves the gain bandwidth, while sacrificing peak gain, peak output power, and power efficiency. However, obtaining the wideband output power requires an additional circuit topology in the reactive matched PA design.

Fig. 7 shows the simulated load pull contour of the 6 × 125 μm GaN HEMTs according to frequencies. As shown in Fig. 7, when the operating frequency increases, the load pull contour rotates in a counterclockwise direction on the Smith chart. Unfortunately, the general output load impedances rotate in the clockwise direction on the Smith chart according to frequencies, which causes the output power characteristics to be very frequency dependent. This prevents the usual PAs from generating uniformly high output power over a wide bandwidth.

We reduce these mismatches between the optimum loads and the output impedances by adopting the L-C resonance technique in the output load matching circuit [15]. First, we set the resonance frequency (in this case, 12 GHz) in the middle of the operating frequency region.

The appropriate value of L and C for resonance is then designed, taking into consideration the length of the micro-strip lines and the size of metal insulator metal (MIM) capacitors. If the capacitance is too large, the power gain will decrease steeply at high frequencies. On the contrary, if the inductance is too large, the layout will be bulky and the MIM capacitors will be too small to be implemented. In this work, the selected values of L and C are 600 pH and 330 fF, respectively. The output load matching circuits are implemented by the distributed L-C components. The distributed L-C components are repeatedly put close behind the output transistors and after combining the two-unit PAs. Micro-strip lines (50 μm wide and 300 μm long) are used as a distributed inductance of about 150 pH. The MIM capacitors (30 μm × 30 μm) are used as a distributed capacitance of about 230 fF. The distributed L-C components bring about a lower Q than is obtained with lumped L-C components [16]. In addition, the use of micro-strip lines of widths, ranging from 20 to 100 μm according to the output load matching section, mitigates the frequency-sensitive output load variation. Both the L-C resonance and distributed components decrease the variation of the output load impedance according to frequencies. Therefore, these methods can bring about a greater improvement in the wideband output power characteristics. The circuit schematic of a two-stage reactive matched PA with the distributed L-C loads is shown in Fig. 8(a). Fig. 8(b) compares the output impedance under the distributed L-C load with the output impedance under the lumped L-C load on the Smith chart. The output impedance under the distributed L-C load passes through a lower Q region that is relatively closer to the optimum loads than it does when a lumped L-C load is applied. Therefore, as shown in Fig. 9, the PAs with the distributed L-C load bring about higher output power than is seen with the lumped L-C load through the wideband frequency region.

The proposed PA MMIC has been fabricated by a commercial 0.25 μm GaN process. Fig. 10(a) and (b) show the circuit schematic and the chip photograph. The proposed GaN PA is designed as a three-stage reactive matched type. The GaN HEMTs (6 × 125 μm) are commonly used in the drive and main power stage. The output stage consists of the distributed L-C output load matching circuits of Fig. 8(a) and a Wilkinson power combiner.

Figs. 11(a) and (b) compare the measured small-signal S-parameters with the simulation. The proposed PA represents the small-signal gain of about 20 dB from 6 to 18 GHz. We presume that errors in the electro-magnetic simulation about the passive components cause partial differences between the simulation and the measurement. Although partial disagreement is evident, the simulation shows reasonable agreement with the measurements from 6 to 18 GHz.

Fig. 12 depicts the measurements of the continuous wave (CW) output power (P_out) and the power-added efficiency (PAE) according to frequencies. Bias voltages supplied at the PA are optimized to obtain the best P_out and PAE. The drastic drop in P_out and PAE at 18 GHz, as shown in Fig. 8, is compensated by slightly changing the final output matching circuit by frequency shifting of the minimum P_out and PAE (13 to 15 GHz). The proposed GaN PA generates the average CW P_out of 5.5 W from 6 to 18 GHz. The measured PAE is about 7.6–23.7% from 6 to 18 GHz. The measured peak P_out and PAE are 39.2 dBm and 23.7%, respectively, at 7 GHz.

The P_out obtained by the simulation using the in-house large-signal models is also compared with the measured P_out in Fig. 12. According to the frequencies, the simulated P_out is underestimated by 0.0–1.7 dB. As in the S-parameter result, the accuracy errors in the electromagnetic simulation with respect to passive components influence the prediction of P_out. The modeled R_th is extracted under the on-wafer condition. In practice, however, the PA is tested under modules in which MMICs are pasted onto a copper jig with gold-tin materials. Thus, the R_th in our GaN HEMT model should be decreased. Compensation of these errors in the PA design should improve the prediction inaccuracy. Some inaccuracy remains, but the proposed large-signal model gives a better prediction of P_out than that provided by the foundry service.

Table 2 compares the performance of reactive matched GaN PA MMICs. The 0.25-μm GaN process used in this work shows a lower F_max, and a poorer load pull power and power efficiency than was obtained with the 0.25-μm GaN processes used in other work [17]. However, the proposed GaN PA shows competitive performance to that obtained in state-of-the-art work. In particular, this work achieves excellent performance in power density when compared to the reported RMPAs.

We implemented 6–18 GHz wideband GaN power amplifiers using a GaN HEMT large-signal model that incorporates a thermal model and is based on various measurements. The inhouse GaN HEMT model is effectively used to design a reactive matched PA up to frequencies close to the device’s F_max. The output load matching circuits are implemented by the distributed L-C components. The output load matching using distributed L-C components decreases the variation in the output load impedance according to frequency. The use of GaN HEMTs with high breakdown voltages and RMPAs with the distributed L-C output load matching successfully achieved an output power over 5 W through the octave bandwidth.