How Algorithms Make Amplifiers for 5G Networks More Efficient?

Amplifiers are the heart of our smartphones' antennas: they turn electrical signals into radio waves. For the new 5G networks we need an amplifier that works over a wide frequency range (1.5‑2.6 GHz) and is highly efficient, otherwise the battery drains quickly. A research team combined two artificial "brains" – genetic algorithms and neural networks – to design a Doherty amplifier, a topology widely used in telecommunications. The genetic algorithm, inspired by natural selection, explores many circuit component combinations, while the neural network, trained on previous results, quickly predicts each configuration's performance. This makes the optimization process faster and more accurate. The outcome is an amplifier that, at full power, delivers between 42 dBm and 45 dBm, with efficiency ranging from 51 % to 62 %. Even at 6 dB back‑off (when the signal is weaker) it keeps efficiency between 44 % and 56 %. These figures show that the hybrid approach can push radio device performance to new limits, helping create faster networks and devices with longer battery life.

The Doherty architecture is a two-stage topology: a main power amplifier (PA) and an auxiliary PA connected through a matching network. At near‑saturation, the main PA supplies most of the power, while the auxiliary PA activates during back‑off regimes, boosting overall efficiency. Traditional optimization of such circuits requires exhaustive search in a highly nonlinear parameter space, which is often impractical. The genetic algorithm (GA) creates an initial population of solutions (sets of values for inductors, capacitors, and transmission lines) and evolves the population through crossover and mutation toward higher‑fitness solutions. The fitness function incorporates multiple objectives: output power (Pout), efficiency (η), and gain (G), weighted according to design priorities. Simultaneously, an artificial neural network (ANN) is trained on results from electromagnetic simulations (e.g., using ADS or HFSS). Once trained, the ANN quickly predicts the performance of new configurations, reducing the number of costly full‑wave simulations. The workflow proceeds as follows: 1) generate a solution pool with GA; 2) perform preliminary evaluation with ANN; 3) select top candidates for full‑wave simulation; 4) update ANN with new data; 5) iterate until convergence. The experimental prototype, covering a 1.1 GHz bandwidth, achieved a saturated output of 44.7 dBm and a peak efficiency of 61.8 % at 2.0 GHz, confirming the hybrid method's ability to balance bandwidth, linearity, and efficiency trade‑offs. This approach scales to higher frequencies (mmWave) and emerging technologies such as digital phase modulators, paving the way for smarter, adaptive RF systems.