In the effort of achieving a low-power and interference adaptive Rx, some previous works have focused on linearizing mixers using frequency translation by compressing third-order intermodulation product (IM3) [4, 5, 6]; however, if the Low-Noise Amplifiers (LNA) are already saturated, linearizing the later stages gains little advantage. Other works contributed on creating ultra-low-power LNA using the current reuse and forward body biasing techniques [7, 8, 9], variable-gain LNA [10, 11, 12, 13], an orthogonally tunable LNA where input third-order intercept point (IIP3) and gain can be individually changed through the bias tuning knobs [14], or a combination of low-power and variable-gain LNA [15]. The variable gain and IIP3 in the design of LNA can assist the receiver to become more interference tolerant in terms of both large signal saturation and small signal nonlinearity; however, these works only included circuit level designs without a complete system design. With the purpose of developing an adaptive receiver, dynamic bias (tuning of the gate and drain voltage) for the optimization of the signal-to-noise and distortion (SNDR) for a Gallium-Nitride (GaN) LNA is explored in [16]; however, the system-level considerations for the dynamic bias technique were not included.

In [17], the authors utilized the orthogonally tunable LNA to implement a use-aware adaptive RF transceiver system where different low-power adaptation modes are designed for different throughput requirements but disregard the tuning speed of the Rx. Other implementations of an adaptive RF communication system use error vector magnitude (EVM) and look up table (LUT) with stored tuning conditions for the LNA and mixer to optimally trade off power and performance [18]. Advancements were made later by considering the process variation of the components with tuning adjustments included in the design process [19]. These contributions provide a thorough design of the Rx system, but most of the results are still simulation-based, and utilization of only the LUT does not provide local feedback for further tuning of the system.

Other adaptive systems with implementation of COTS, IC, or simulation can be found in [20, 21, 22, 23, 24, 25, 26, 27].

This paper presents the first building block, a high input tolerant and interference adaptive GaN LNA front-end, to an instinctual adaptive receiver. Fig. 1 modifies the traditional RF receiver with the instinctual interference adaptation by incorporating 1) sensing through the onboard envelope detectors as observation points at the input and/or output of the GaN LNA, 2) processing through localized digital processing unit or the auxiliary ADC that’s already present in baseband(BB), and 3) feedback to the customized GaN LNA bias controls for the best linearity and power performance while maintaining signal integrity. Note that in this paper, the processing is done using an MCU, but it can also be done using the BB DSP and auxiliary ADC. When blockers are present in a traditional Rx system working in nominal conditions, the blockers would saturate the LNA and produce a comparable IM3 to the actual signal which results in an undecidable LNA output. When blockers are present in the front-end with instinctual interference adaptation, the control logic would be able to increase the linearity of the system which in turn increases the IM3 compression at the cost of power consumption. When the signal and interference are both low, the Rx with bias control would be consuming less power for approximately the same signal levels. Note that the need for a better linearity range is present regardless of whether the high power is from the desired signal or the interference as long as the system is able to be brought back to the linearity range.

To have a higher power handling capability and linearity, GaN LNA is utilized. Our prior work [1] involves interference adaptation for an Rx system with incremental adaptation control involving both feedforward and feedback path for the GaN LNA. However, the utilization of bench-top equipment significantly increases the adaptation time and the form factor of the system, which makes the design unsuitable in real life. [1] also lacks the consideration of the effects of the GaN LNA properties.

This paper builds upon the prior work and has the following additional contributions:

• •

  This work implements the first sub-1ms interference adaptive, instinctual GaN LNA system with a localized in-built intelligence using a microcontroller. The system consumes $\leq$10$\%$ of nominal LNA power to provide a wide tuning range of linearity for about 11 dB and LNA power for 0.5-2 W.

• •

  The control circuitry of GaN LNA has been designed with careful consideration of the high-power effects of GaN LNA and the trade-off between system adaptation time and device lifetime.

• •

  Background control theory of the system is provided on the limitations for the overall adaptation time (< 1 ms) which shows a high correlation with the measurement results. To the authors’ current knowledge, this is the first control theory introduced for an interference adaptive RF front-end system.

• •

  Important trade-offs are presented between two designs: 1) feedforward + feedback using control theory mentioned in Sec. IV-E and 2) feedback only using i) incremental adaptation, ii) lookup table (LUT) and iii) one-shot + incremental adaptations. The different designs illustrate different timing constraints and complexities for different applications.

The paper is organized as follows: Sec. II provides an overview of the control loops and component characterization; Sec. III investigates different design considerations such as the design of directional coupler, high input power effects for a GaN LNA, Gate Voltage (V_G) based tuning method and design comparison; Sec. IV describes the three control methods and the control theory; Sec. V presents the measurement results and Sec. VI presents the future directions of this work.

The two-layer PCB utilizes Rogers 4003C material with a thickness of 0.508 mm. The signal lines are carefully matched to 50 $\Omega$. The system architectures are shown in Fig. 2 with associated board layouts shown in Fig. 3. The commercial off-the-shelf parts (COTS) used on the PCBs are listed in Table 1. Fig. 2a) and 3a) show the architecture and layout for the feedback-only design where the interference detection and adaptation are performed in the feedback loop that utilizes an envelope detector (ED) to sample the output of the LNA. The bias control circuitry for the LNA is shown in Fig. 2b). Fig. 2c) and 3b) show the architecture and layout for the feedforward and feedback design. The feedforward loop involves the use of envelope detector 1 (ED1) that’s higher in sensitivity than envelope detector 2 (ED2) used in the feedback loop while still able to sustain the maximum input signal limited by the rating of the LNA at 30 dBm (1W). The difference between the two designs is how the interference will be detected (through the feedforward and feedback path, or directly from the feedback path) which will be discussed more in later sections.

The directional couplers and the LNA forms the front end in the system that will be connected to a mixer and baseband stages in a standard RF receiver. For a combined feedforward and feedback system, RF input first passes the through port of the directional coupler 1 to the LNA while the coupling port enables ED1 to measure the input signal level. Directional couplers are necessary for measurement because of the physical constraints (power handling capability) of the EDs as well as to avoid a significant power divide from the LNA. If the input signal is high (i.e. a blocker), then ED1 will output a higher DC voltage. The assumption is that high-level input signals are only caused by the interference and not the desired signal, thus high input signal corresponds to a high interference level. ED2 in the feedback path will also be incorporated in the interference detection for a better sensitivity after the LNA gain. The measurement will be processed through the ADC on the microcontroller and if the microcontroller determines the adaptation of the LNA is needed, tuning control will be initiated. The feedback control adjusts the gate voltage (V_G) of the LNA to achieve the high linearity for the LNA which will be further discussed in Sec. IV. Note that the drain voltage (V_D) of the LNA is not being controlled because the linearity improvement is minimal with changing V_D for the specific LNA presented; however, V_D controllability can be considered for future improvements on other LNAs. The difference with a feedback-only system is that the presence of the interference will only be detected using ED2 in the feedback loop.

The components used are described in Table I. The LNAs in both the MWCL [1] and this paper are the same part number, but there are some chip-to-chip variations such that with the same V_G, I_D is higher in this paper. Note that the characteristics of the GaN LNA are determined more by the current rather than the bias voltage. This is because diodes are important in the GaN LNA modeling, and biasing the diode current with the correct voltage is more important [36]. We decided to continue with the V_G range of -2.7 V to -2.2 V with a higher I_D due to a worse S11 response at V_G < -2.7 V as shown in Sec.V-C. The feedforward and feedback path components are characterized in the 2-6 GHz range. The losses due to the SMA cables are calibrated during the characterization of the RF components and measurements thereafter. RF components are chosen to have an input impedance of 50 $\Omega$. Fig. 4a)-c) represent the behavior of a GaN LNA at 4 GHz. As LNA’s V_G increases from -2.8 V to -2.2 V, the gain first increases and saturates at around -2.5 V, then starts to decrease. The gain also increases with increasing V_D. Because the change in gain is less than 4 dB (neglecting the -2.8 V V_G data due to low gain and P_1dB,IN, the adaptation will be between V_G=-2.7 V to -2.1 V), the LNA gain is a weak function of both V_D and V_G. The supply current (I_D) increases by more than a hundred mA with increasing V_G while having a small change with increasing V_D. This makes LNA’s I_D a strong function of V_G and a weak function of V_D. The input P_1dB (P_1dB,IN) of the LNA varies around 16.5 dBm with increasing V_G but changes minimally with V_D. This makes the LNA’s P_1dB,IN a strong function of V_G and a weak function of V_D. Therefore, only V_G will be changed in the later sections to achieve linearity while V_D is fixed at 10 V. Note that if the LNA’s P_1dB,IN response is both a strong function of V_G and V_D, or if the LNA’s gain is a strong function of V_D, both V_G and V_D can be implemented in the control loop.

Fig. 4d) represents the behavior of ED2 at frequencies of 4, 5, and 6 GHz. The voltage output of ED2 increases exponentially with increasing power. ED1 is chosen to be able to detect low power levels at the input. Other controlling components listed in Table I are generally chosen to be low power, short settling and rising time, and low noise.

In our prior work [1], we utilized COTS directional coupler components to provide the measurement path; however, even though the components have outstanding specs, the integration of the directional coupler with the PCB board is causing unwanted reflections from the soldering and the abrupt transition from the trace to the component. Consequently, this paper takes advantage of the onboard microstrip design for the directional coupler which avoids extra transition from the board to the component. The coupler is necessary for decreasing the power input to below the power limit of the envelope detector and not to diverge extra power from the signal path for measurement purposes. The microstrip directional coupler schematic is shown in Fig. 5a) [37]. The measurements for the directional coupler are shown in 5b)-d). The S11 shows low return loss with measurements below -19 dB over the frequency from 2 GHz to 6GHz. The insertion loss (S21) increases from 0.18 dB to 0.42 dB with frequency. The coupling (S31) changes from 20.5 dB to 16 dB over the bandwidth.

Unlike the typical GaAs LNA’s lifetime being limited by the breakdown voltage, GaN LNA is limited by high DC gate current and drain-gate voltage (V_DG). Characteristics of a GaN LNA with high input power have been investigated in [38]. Similar characteristics have been presented for the LNA used in this paper in Fig. 6.

Fig. 6 describes I_G, V_G and I_D vs. different input power with different resistor values in series (R_S) at the gate bias of the LNA with the set up in Fig. 6a). The frequency has been set at 2 GHz for the worst-case scenario. As shown in the plots 6b)-d), low power characteristics are very different than the high power characteristics. In the low power region or the adaptation range, I_G is actually negative causing V_G to tail up due to R_S, and with Ohm’s law, when I_G is flowing out of the gate, V_G is higher than the control voltage (V_C). Because of the higher bias in V_G, I_D is also drawing more current. With increasing in R_S, V_G tails up higher and I_D also reaches a higher point. When the input power transitions to the high power region, I_G becomes positive and starts to increase exponentially. As a result of I_G becoming more positive, V_G starts to drop off exponentially with I_D also dropping. With increasing in R_S, I_G increases more gradually which protects the LNA. Even though V_G drops to about -10 V range, it is not significant to cause the LNA to break down. The typical critical value for V_DG causing degradation in LNA performance is 30 V, where if V_D is 10 V (used in this paper), the minimum V_G is -20 V [39].

The characteristics of I_G with varying P_IN is explained by the Shockley contact at the gate and source of a GaN high-electron-mobility transistor (HEMT). A GaN LNA model can be found in [36]. The Shockley contact is essentially a diode, consequently, at low power levels, the gate experiences a small leakage current; at high power levels, the diode is turned on and I_G increases exponentially [40].

Another effect of the GaN LNA is the trapping effect on the gain recovery after a pulse of high input signal [3]. This effect happens much higher than the adaptation range (P_IN > 25 dBm), thus during the pulse, the system can be tuned at the maximum possible linearity. When the high input signal is removed, the slow gain recovery does not affect the detection that the input signal is now minimal and returns back to the low power mode.

To protect the GaN LNA against the high input power characteristics and provide reasonable adaptation time, the current design for tuning the V_G of the LNA is proposed in Fig. 7c). The microcontroller will be configuring the digitally programmable potentiometer (DPP) which is supplied by a -5 V inverting charge pump for the negative V_G bias. Following the DPP is a buffer and a parallel structure of a resistor and switch for the control of I_G.

The process for the current design in tuning the V_G is shown in Fig. 7. Fig. 7a) considers the case when there is only a DPP connected directly to the gate of the LNA. The drawback of this design is shown in Fig. 8. As the R_S to the gate of the LNA increases, the gate current is suppressed but the settling time increases. Since the DPP is in the 100 k$\Omega$ range, the high series resistance causes a high RC time constant which increases the settling time (T_S) significantly. The increase in Ts would, in turn, increase the adaptation time to tens of milliseconds which is undesirable when one of the goals is to constraint the timing in the millisecond range. To solve the Ts problem, a buffer is added as shown in Fig. 7b) which essentially reduces R_S to decrease T_S; however, as discussed in Sec. III-B, the reduced R_S draws more I_G in high power which minimizes the device lifetime.

To balance the I_G and Ts trade-off, a parallel resistor-switch pair is added in the current design as shown in Fig. 7c). During the low power and adaptation regions in Fig. 6a), I_G is low in the sub-mA region, so the switch can be closed to construct the low resistance path for tuning V_G to improve Ts. After the adaptation, the switch will remain closed to maintain a steady bias condition, on the other hand, if the switch opens to place R_S in the path, V_G would increase as shown in Fig. 7b) and change the bias condition which consumes more power. In the high power region, lower I_G is more prominent since V_G is set to -2.1 V directly without the adaptation, so T_S can be traded for lower I_G with the switch being open and resistor in series.

Due to the variable gain behavior of the LNA with changing V_G, the ED2 measurements at the output of the LNA are also affected. The gain of the LNA first increases with increasing V_G, then decreases after V_G$\approx$-2.5 V in Fig. 4a). Fig. 9 shows that the linear region occurs where the gain of the system remains relatively constant before the 1dB compression point (P_1dB) and the nonlinear region occurs where the system is highly compressed beyond P_1dB. Even though the gain varies with increasing V_G, P_1dB monotonically increases with increasing V_G. When a high input signal presents in the system at V_G=-2.7 V, the signal is highly gain compressed and the system is highly nonlinear. To linearize the system, V_G increases in 0.1 V increments. When V_G increases from -2.7 V to -2.6 V, the signal becomes less gain compressed but the system remains nonlinear. To further improve the system linearity, V_G is increased to -2.5 V. The idea of incrementing V_G to improve linearity forms the basis of the incremental adaptation in Sec. IV-A.

Fig. 10 shows P_OUT vs V_G tuning with respect to different P_IN. P_1dBs are also shown to show that V_G < V_P1dB results in a nonlinear system. As P_IN increases, the V_P1dB also increases. As V_G is tuning, the gain of the LNA increases then decreases as shown in Fig. 4a). Due to the effect of LNA gain with different VG and LNA transitioning from non-linear to linear, the determining step decreases with P_IN increase. Therefore to accommodate the difference in determining steps in both high and low P_IN across different frequencies, a lower determining step is chosen as the linearity threshold which has a drawback of overestimation of V_G for lower P_IN values. One may suggest that different threshold limits can be used with different P_IN values, but this again creates a case-dependent threshold that may not work in other frequencies.

Fig. 11a) describes the control method with incremental adaptation. The LNA is initially in low power mode with V_G being initially set at -2.7 V and ED2 will perform the initial measurement at time instance 1 (t1) as the reference. When interference reaches the system, another ED2 measurement at time instance 2 (t2) is taken as the current value. The difference (err) between the current and reference of ED2 is compared with a set of thresholds to start the feedback control. The set of thresholds can be regarded as an extended version of bang-bang control as the triple set-point control in Fig. 11b). The triple set-point has three different conditions: when V_G1 at t1 is the same as V_G2 at t2, the resulting action is increment V_G by 0.1 V if the err is greater than the positive threshold (+Th), decrement V_G by 0.1 V if the err is less than the negative threshold (-Th), or maintain the same V_G if the err is between +Th and -Th; when V_G2 is greater than V_G1 (VG is incrementing), the err is compared only to the +Th such that V_G incrementes when err is greater than +Th (gain compression observed) or returns to V_G1 when err is less than +Th (linear); when V_G2 is less than V_G1 (VG is decrementing), the err is compared with the -Th and a more negative threshold (–Th) such that V_G decrements to find the optimum V_G when the err is in between the thresholds (still in the linear region), or returns to V_G1 when the err is less than the –Th (gain compression observed). An example of the incremental adaptation is shown in Fig. 11c). When the interference is detected, V_G will start incrementing until err is less than the +Th. Another example of when the interference signal changes are shown in 11d). When the interference level decreases, the V_G also starts to decrement until err is less than –Th. Note that if the interference level drops a significant amount, V_G is back to the initial condition instead of stepping down to start the adaptation.

Fig. 12a) describes the control method with a look-up table (LUT). A LUT as shown in Fig. 12b) utilizes the current V_G value and the ADC measurement (unit: 10^-4 V) of the ED2 at 3 GHz, then outputs the V_G value that would bring the LNA back to linearity (High linearity mode). Note that, although not included in the figure, after settling of V_G, ED2 is constantly monitored so that if the measurement is within the ADC variation, V_G stays at the same value; otherwise, the process starts over with another ED2 measurement. The LUT is generated in a way that for each input power using the P_1dB,IN at each V_G value and different V_G settings, ADC measures at the output of ED2. LUT can be associated with a Fuzzy logic control that unlike the triple set-point control to only have four commands, Fuzzy logic includes a wider range of V_G outputs. Each combination input V_G and ED2 measurement can be treated as an if-else statement in the Fuzzy logic and returns a preprogrammed output V_G [41]. An example of the LUT control is shown in 12c). When the interference is detected, ADC measures at ED2 output. With the ADC measurement of 743.8 mV (ADC reads 7438) and the current V_G value of -2.7 V, the LUT determines that a V_G value of -2.5 V with a P_1dB,IN of -6 dBm is sufficient to bring the system back to linearity. Another example is shown in 12d) in the case of reduced interference level. Again the ADC measurement and current V_G value is used to determine that a V_G of -2.6 V with a P_1dB,IN of -8.5 dBm is sufficient for linearity.

Fig. 13a) describes the control method with a one-shot for rough tuning and incremental adaptation for fine-tuning. The one-shot is implemented in a LUT style with certain degrees of an underestimate of the output V_G value to accommodate different frequencies and different V_G values. The incremental adaption again utilizes the triple set-point control with three thresholds and four regions of action. An example is shown in 13b) where when the interference of -5 dBm is detected, one-shot rough tunes the V_G to -2.5 V, then the incremental adaptation starts to increment V_G for fine-tuning. A second example is shown in 13c) where when the interference level drops significantly, one-shot tunes V_G to -2.7 V, and then incremental adaptation steps up to find the optimum V_G.

All three methods are intended to control the bias of LNA in order to achieve linearity with a minimum required power consumption from the LNA. LUT has the advantage of being very fast and accurate if the frequency is known so that the specific LUT can be utilized; however, accuracy implies a stringent requirement on the memory space in the microcontroller that multiple LUTs at different frequencies are required. On the other hand, the incremental adaptation only and the one-shot + incremental adaptation methods do not need the frequency information for adaptation as long as the measurement is greater than the interference threshold. However, due to the extra adaptation and the single set linearity threshold, the two methods tend to take longer and are less accurate involving some overestimation of V_G. A way to incorporate methods is that when the interference frequency is known, LUT can be used, but when the frequency is unknown, one-shot + incremental adaptation can be used.

In order to provide a more thorough background on the limitations of the control loop adaptation time, a second-order dynamic system is presented in Fig. 14 following analysis for digital dropout regulators [42, 43]. The control loop models the feedforward and feedback control method with two degrees of observability at the input and output. The model has the following assumptions:

1. 1.

   flat gain until the P_1dB point.

2. 2.

   gain is not a function of V_G and frequency but linearity.

3. 3.

   V_G settling (100kHz) dominates ED bandwidth (40 MHz) and LNA bandwidth.

The goal of the control loop is to minimize the difference between the expected output and the measured output. The expected output is calculated from the measured input and multiplied by a constant LNA gain (G). Both the expected output and the measured output are sampled and subtracted to form an error signal. The error signal multiplies with a proportional constant (K_VG) to form a $\Delta$ V_G that is added to the previous V_G through the integrator ($\frac{z}{z-1}$). The sampled V_G transforms to continuous time through the zero-order hold. Finally, V_G supplies to the plant and maps the V_G to the measured output. The plant consists of a constant gain K_G and a pole from the V_G settling time (F_Load) with a z-domain equation of $\frac{K\textsubscript{G}(1-e^{(-F_{Load}T_{s})})}{z-e^{(-F_{Load}T_{s})}}$. T_s is the ADC sampling period(we will approximate to 50$\mu$s for simpler demonstration of calculation). The open loop gain is

To maximally match the existing control methods, since each adaptation step requires two samples of measurement to ensure accuracy which will be explained shortly, 2*T_s=100 $\mu$s with a sampling frequency F_s/2=10 kHz is used. With F_Load=100 kHz, F_s < F_Load, F_s dominants in the settling time. From Fig. 15a)-b), the proportional constant is low with K_VGK_G=0.4, the total settling time or the adaptation time is 771 $\mu$s. From Fig. 15c)-d), K_VGK_G=0.6, the total settling time is 436 $\mu$s. As shown in the step response, because T_Load < T_s, each step takes 2*T_s. The different proportional constant can be thought of as the one-shot values in the one-shot + incremental adaptation method, with a higher one-shot value (high K_VGK_G), the number of steps to reach a steady state is lower, hence a faster response.

Overall, the adaptation time can be approximated as

where

In the equations, N is the number of steps which is a function of the proportion constant, T_s is the ADC sampling time, T_process is the processing time for the microcontroller to make decisions, T_VG is the V_G tuning time, T_LNA is the propagation time from tuned V_G to the settling of the LNA and T_ED is the propagation time from settled LNA output to settled ED output. Note that the first sample of the ADC contains some of the transient response which lacks accuracy, so the second sample is taken to be the accurate one to the processing, hence the 2*T_s in Eq.2. In order for the second sample to be accurate, the first sample of the ADC should contain all of the transient responses, hence Eq.3 shows that the sampling time of the ADC needs to be greater than the total propagation and settling time for each component.

Some known timing characteristics are:

• •

  T_VG $\approx$4 - 10 $\mu$s per step depending on the step size

• •

  T_process is negligible in the ranges of <$\mu$s

• •

  T_s$\approx$2 $\mu$s - 1 ms or F_s$\approx$1 - 500 kHz

• •

  T_Load$\approx$ T_VG$\approx$4 - 10 $\mu$s or F_Load$\approx$ F_VG$\approx$100 - 250 kHz.

Since highest F_s is 500 kHz, F_s << F_Load, a linear increase of adaptation time vs. ADC sampling time is observed in Fig. 16, assuming T_VG$\approx$ 10$\mu$s and a negligible processing time with K_VGK_G=0.6. Note that in the figure, a sufficient margin higher than T_VG ( >25us of total ADC measurement time for 2 samples) is needed for an accurate reading. The simulations in Fig. 16 allow for a more accurate estimation of the minimum adaptation time for the control loop. The adaptation time is strongly dependent on the ADC sampling rate, so if ADC sampling rate can be increased, the adaptation time can be reduced. Otherwise, the minimum adaption time is about 150 $\mu$s with the current sampling rate with the number of steps to settling being approximately four.

Feedback-only control implements a temporal control where the tuning uses two samples of the output signal levels at different times. The results are summarized in Table II.

Fig. 18 presents the adaptation and timing characteristics for different control methods with the measurement setup in Fig. 17. Note that after zooming into the transient, the RF output measured using an oscilloscope is not a perfect sine wave due to the undersampling of the oscilloscope trying to capture the signal with a larger time scale. Fig. 18a) - c) show the adaptation of the LNA by varying different interference levels on and off. Different high interference levels at 3 GHz are forced at the input with the expectation of V_G to increase from -2.6 V to -2.1 V in the increment of 0.1 V; at low interference levels, V_G is expected to drop down to -2.7 V. A summary of different settled V_G with respect to different input interference levels is presented in Fig. 19a). As expected, the LUT(Fig. 18b) is more aligned with the expected V_G as the LUT values are specifically for 3 GHz while both the incremental adaptation only(Fig. 18a) and the one-shot + incremental adaptation(Fig. 18c) overestimates V_G. Fig. 18a1), b1) and c1) show the tuning times to adapt to an interference level for the incremental adaptation, LUT and one-shot + incremental adaptation are 580, 180 and 450 $\mu$s respectively. Fig. 18a2), b2) and c2) show the tuning times to adapt to a disappearing interference for the three control methods are very similar, around 170 $\mu$s. Fig. Fig. 18d) - f) show the adaptation of the LNA when the interference level increases from -12.5 dBm to -0.5 dBm and backs down to -12.5 dBm in increments of 1 dBm. With the LUT method in Fig. 18e), every step of V_G is shown and roughly the same settling V_G for the same interference level, whereas for the incremental adaptation (Fig. 18d) and one-shot + incremental adaptation method(Fig. 18f), some V_G steps are skipped and sometimes different V_G values for the same interference level. Fig. 18d1) shows that when the interference level decreases, the control loop is able to step down and adapt.

Fig. 19 shows the settling V_G values for different methods of control vs. input power with frequencies of 3 GHz and 2.5 GHz. In the 3 GHz plot in Fig. 19a), LUT method almost matches up with all the expected V_G to bring the LNA back to linearity whereas the incremental adaptation only and one-shot + incremental adaptation overestimates for many P_IN values. However, in the 2.5 GHz plot in Fig. 19b), LUT significantly underestimates the necessary V_G for linearity due to the LUT is only captured at 3 GHz and the algorithm tries to match the 2.5 GHz ADC values with the 3 GHz values. On the contrary, the one-shot + incremental adaptation method matches up with the expected better at 2.5 GHz than at 3 GHz due to the set threshold value being better suited at 2.5 GHz since the threshold value is chosen to adapt to a wider range of frequencies.

Fig. 20 shows the trade-offs between an adaptive system vs. a non-adaptive system operating in nominal conditions with a two-tone measurement. Fig. 20a) and b) show that when the interference is low, the adaptive system consumes less power while still maintaining linearity for the LNA in comparison to the non-adaptive system having a higher IM3 compression and consumes more power; when the interference is high, the adaptive system consumes more power to bring the LNA back to linearity with a higher IM3 compression compared to the non-adaptive system with lower IM3 compression (LNA is non-linear) and lower power consumption. Fig. 20c) shows that the adaptive system is able to keep the LNA in the linearity range over a wider input range than the non-adaptive system in the nominal condition.

Feedforward + feedback design allows a spacial control of the system where both the input and output signal levels can be sampled, and the linearity is determined by comparing the measured gain with the expected gain to ensure no gain compression. Fig. 21 shows that the tuning circuit is able to use spatial control to tune the V_G of the LNA as the interferences appear, increase and disappear; however, this method also suffers from an overestimation of the V_G value that a higher V_G value is determined. Note from Fig. 21a1), unlike the overshoot in the feedback only incremental adaptation control, since this board has the extra degree of observability and that the measured gain is directly compared with the expected gain, tuning does not give overshoot in V_G. From Fig. 21a1-3), the tuning times for the interference appearance, increasing of interference, and disappearance of interference are 770 $\mu$s, 140 $\mu$s and 250 $\mu$s respectively. The tuning times are below 1 ms; however, as shown in Fig. 21a1), a longer tuning time is needed than the feedback only incremental adaptation control due to ADC sampling time is lengthened to accurately measure the input power being close to the sensitivity level.

Fig. 22 presents different LNA characteristics after implementation on the feedback-only board and the combined feedforward and feedback board with the datasheet values in similar bias conditions of V_D = 10 V and I_D $\approx$ 100 mA. Note that the datasheet values are measured directly on the die, by implementing on a PCB, some degree of degradation in the gain and NF are expected. Fig. 22a) shows the gain comparison over the frequency range of 2-6 GHz. The gain of the boards shows a few dB of degradation in comparison to the datasheet while little difference between the different boards. Fig. 22b) shows the noise figure (NF) vs. frequency for different boards. The feedback-only board has an NF about 0.5 dB higher than the datasheet, and as expected with the feedforward and feedback combined board, the NF is again about 0.3 dB higher than the feedback-only layout due to the extra directional coupler before the LNA. Fig. 22c) shows the 3rd order intermodular product (IM3) compression vs. desired output power for each tone. The IM3 compression are relatively close across different boards and the datasheet.

Fig. 22d) shows different return loss (S11) with respect to different V_G values. As shown in the plot, when V_G=-2.8 V, the return loss is worse than all of the other voltages, so when implementing the control loop, only V_G greater than -2.8 V is considered. For V_G between -2.7 V to -2.1 V, the majority of the responses have an S11 lower than -15 dBm, other parts have an S11 lower than -10 dBm. Fig. 22e) shows NF across V_G tuning range for 2, 4 and 6 GHz. Across V_G the NF can increase about 0.4 dB. When interference is low, instead of operating at nominal V_G of -2.4 V with a higher NF, lower NF can be achieved in the adaptive system with V_G of -2.7 V. When interference is high, higher linearity is achieved with higher NF and power consumption.

Fig. 22f)-h) shows the comparison of the high power data for the LNA in Fig. 6 and LNA in a feedback system. In the adaptation range, the V_G in the feedback system is continuously increasing to maintain linearity by adapting to the current input while the V_G increase in the LNA only system is caused by the non-linearity of the LNA. The difference in linearity in the two systems can also be observed in the I_G, where the linear system has a consistent I_G in the ranges of $\mu$A whereas the non-linear system sees more of a revered biased current. I_D for both systems show an increase because of the increase in V_G. In the high power range, the currents and the voltages are starting to overlap showing similar high power effects as described in Sec. III-B.

Table III shows comparisons between the datasheet LNA in nominal condition, LNA with feedback control and LNA with feedback + feedforward control. According to the datasheet [28], the GaN LNA’s nominal condition is V_D=10 V and I_D=100 mA, which consumes about 1 W of power. The controls consume $\leq$10$\%$ of total LNA nominal power for a better linearity and LNA power consumption trade-off. Most of the control power derives from the microcontroller consuming 80 mW of power, which is about 80-90$\%$ of total control power. LNA with control gives a wide tuning range in power and linearity. When the system is operating in low-power mode in a large EW receiver array, the power saving can be significant. Extra system power is consumed when higher linearity is required to decode the signal.

Feedforward + feedback design includes two pairs of directional coupler and ED. With the extra directional coupler and ED1 before the LNA, NF increases by $\approx$ 0.3 dB, and overall control power consumption increases by another 10 mW than feedback only design; however, the design gives more information on the input signal that will be more relevant in future works of an adaptive receiver. For example, if a filter was placed to remove the interference signal, the extra information on the input signal would allow us to determine if the interference is still present or has already been removed by the filter. The feedback-only design provides a simpler solution for the current application to detect the presence and the level of interference signal at the output of the LNA. Contrary to the spacial control in the feedforward + feedback design, feedback-only design implements a temporal control. At the current stage, the extra degree of observability at the input of the LNA is not needed as no filter is present.

Table IV shows the different works that contribute to adaptive transceivers over different implementation methods for higher resilience towards undesired interference.