Artificial intelligence (AI) and machine learning (ML) technologies are pervasive in our daily life empowering devices ranging from smart speakers to thermostats, self-driving cars to robots and social networks to banking systems. In wireless communications, ML has been recently applied across all layers including network planning, spectrum sensing, channel modeling, security and even the smart applications running on our mobile devices. Meanwhile, some are envisioning a future communication system that brings the hyper-connected experience to every corner of life in beyond 5G and 6G.1 Application and deployment of AI technology for next-generation wireless communications have the profound potential to improve the end-to-end experience and reduce both the CAPEX and OPEX of networks.2 AI becomes a necessary tool for delivering reliable and versatile services to connect hundreds of billions of machines and humans.
Improving radio hardware performance of radio access network, particularly, RF power amplifiers (PAs), has been a long-lasting challenge with ever-increasing system demands. In the past decades, RF engineers have spent numerous efforts to enhance PAs figure of merits such as power efficiency, gain, bandwidth and linearity. They came up with many brilliant solutions. Nevertheless, as the complexities of advanced PA circuits, modules and systems keep increasing, it becomes even more challenging and time consuming to design, operate and optimize PAs for highly dynamic signals with fast varying envelopes, dynamic network traffic and beam dependent radio environments such as massive-MIMO. However, such challenging use cases are becoming very common for modern mobile communications.
This article focuses on the recent studies of introducing ML for radio frequency PAs’ online operational conditions optimization, primarily at sub-6 GHz frequency of 5G. Two demonstrators of advanced PA architectures are designed with cutting edge 0.15 μm GaN high electron mobility transistor (HEMT) technology, namely: a digital Doherty power amplifier (DDPA) and an innovative digitally assisted ultra-wideband mixed mode dual-input PA based on frequency-periodic load modulation (FPLM). For both examples, compact data-driven ML techniques are applied to significantly boost PAs performance. Combined with innovative hardware design, AI and ML can become a powerful tool to assist RF engineers dealing with sophisticated PAs design and operating challenges.
DIGITAL TO INTELLIGENT DOHERTY PA
Doherty PAs have been the workhorse for cellular base station radio transmitters3 thanks to its relatively simple topology and attractive average power efficiency for amplifying signals with high peak-to-average power ratio (PAPR > 6 dB). Due to its active load pulling principles and analog nature, Doherty PAs still suffer from several key limitations such as non-optimal power splitting ratio, phase alignment and peaking amplifier turning-ON, especially over wide RF bands and input power levels.
Figure 1 A wideband Doherty power amplifier8 and its modified version for dual-input Doherty power amplifier. The modified input power divider is outlined.
To overcome such difficulties, various modified design methods and architectures including Advanced Doherty Alignment module and DDPA were proposed by eliminating the conventional analog-based power splitting circuitry (i.e., Wilkinson divider). Instead, these designs are feeding dual-input RF signals directly to the Carrier and Peaking amplifiers of the Doherty PA, respectively.4,5 Hence, the circuit can independently control input signals amplitudes and phases with better results. Figure 1 provides a comparison of a Doherty PA and its modified version as dual-input DDPA. The input network change is highlighted in the figure.
Multi-input Doherty PA can be digitally controlled by following a set of derived closed-form equations, which approximate a pre-determined static power splitting ratio and phase imbalance between Carrier and Peaking amplifiers. Alternatively, it can be done by offline brute force search, finding an optimum input signal condition for high efficiency or high output power.5-7 However, these two approaches have several limitations in practice: (1) derived mathematical equations only provide an approximation of highly non-linear relationship within PA (i.e., using arctan function), (2) bias voltages optimization is not included but critical and (3) open-loop implementation does not capture the device-to-device variation or operating condition changes (i.e., ambient temperature). Consequently, manual tuning is still required to account for the dynamics of real systems and condition variations. Because of the large searching space of variables, brute force searching is inefficient for practical implementations.
Very recently, there have been new ML data-driven online optimization methods proposed and demonstrated. In an initial study shown by simulation,9 a Simultaneous Perturbation Stochastic Approximation (SPSA) algorithm was applied to optimize the input power splitting ratio, phase offset and gate bias voltages at the same time for the Carrier and Peaking amplifiers of a dual-input Doherty PA using ADS and SystemVue software. The algorithm is here:
Figure 2 Online optimization of dual-input Doherty PA.
It formulates digital DDPA real-time optimization as an adaptive online control problem by searching for an optimum solution for a user defined cost function consisting of several PAs figure of merits (a weighted sum of power, gain, efficiency and linearity etc.), as depicted by Figure 2. Different hyper-parameters and initial conditions of optimization were tested. As a result, optimal points of power added efficiency can be found between 60 to 70 percent with many closely spaced local minimum points. A further development with a lab test bench, shown in Figure 3, is a proof-of-concept and engineering demonstration that was implemented.10
Figure 3 Testbench of dual-input digital Doherty PA with machine learning online optimization.
One setup implemented a model-free optimization method with simulated annealing (SA) and extremum seeking (ES), as shown in Figure 4.10 The combination of SA and ES makes the system optimization efficient. SA captures the random and abrupt variation in the system mainly due to frequency and input power level variations, where ES captures slow variation in the model such as temperature.
Figure 4 Model-free machine learning algorithms used for DDPA optimization.10
Figure 5 DDPA performance with online auto-tuning of control parameter: Pout and defined cost function (a), gate bias voltages for main and peaking amplifier (b), PAE and Gain (c) and input phase imbalance and power splitting ratio (d).
The compactness of ML algorithm adopted here is Quite different from the general deep learning ML category, such as deep neural network, in the sense that it neither requires massive training data nor powerful computation power and memory. This is an important feature for efficient implementation of RF front-end applications. Figure 5 shows DDPA online auto-tuning of performance including output power, gain, power added efficiency via adaptive control of gate bias voltages (Vg_main, Vg_peak) and input power splitting ratio (α: how much power distributed to Peaking amplifier from total input) and phase imbalance (ΔΦ) using SA and ES. The optimization goal is to search for an optimal control parameters θ* maximizing cost function Q(θ), which is expressed as the weighted sum of PA performance of interest: θ* = argmaxQ(θ), θ∈U, where θ is a vector of the amplifier tuning parameters defined as θ = [Vg_main, Vg_peak, ΔΦ, α].
As shown in Figure 5, it takes approximately 40 iterations for SA to perform random exploration with Quick convergence, limited mainly by the interface communication of the test instruments. SA is then followed by ES algorithm for a fine tuning to account for effects such as temperature changes. The program is written in MATLAB and running on a PC controlling the measurement setup depicted in Figure 3. Significant performance enhancement in DDPA over a wider frequency range and different input power range (in particular lower input power range) has been observed compared with single input conventional Doherty PA thanks to the auto-tuning procedure. Over a 15 percent efficiency boost and 2 to 3 dB gain is realized without using digital predistortion (DPD). The algorithm is also able to figure out a reasonable tradeoff among these conflicting PA performance targets by assigning different weights in Q(θ). It must be mentioned that dedicated DPD schemes were not used.10
