← All projects

Research walkthrough · Optimization

Power amplifier design

BNN-assisted layout optimization with EM and harmonic-balance validation.

2 min read

Read the paper

E-GASPAD combines local Bayesian neural network surrogates with a hybrid evolutionary search for layout-level power-amplifier design. The aim is to satisfy a complete set of RF specifications using costly EM and harmonic-balance evaluations, without requiring a high-quality initial design.

Method

  1. 01Parameterized layoutEM + harmonic-balance evaluations
  2. 02Local and global searchDifferential-evolution operators
  3. 03Local BNN modelsUncertainty-aware prescreening
  4. 04Validate and diversifySimulate + reconstruct the population
↻ Local refinement improves candidates; population reconstruction preserves search diversity. Method summary redrawn from the paper.

Local BNN models screen candidate designs. Local refinement improves promising solutions; a global phase and population reconstruction maintain diversity. The implementation connects MATLAB optimization to ADS / Momentum using parameterized subcircuits and AEL scripts.

A wideband Doherty layout

Doherty power-amplifier layout with named design variables
The 24–31 GHz Doherty MMIC layout. The study optimizes 31 design variables against 10 specifications. Paper Fig. 9 ↗

Convergence and RF performance

Average convergence of E-GASPAD and GASPAD for the class-AB case
Average convergence across four runs in the class-AB example; the objective measures specification violation. Paper Fig. 7 ↗
Case Variables / specifications Mean full evaluations Mean elapsed time
27–31 GHz class-AB 27 / 7 516 52 h
24–31 GHz Doherty 31 / 10 574 60 h

The means are reported over four independent runs per case.

Simulated gain and power-added efficiency of the optimized Doherty design
Simulated large-signal gain and power-added efficiency across the Doherty operating band. The PA validation in this paper is simulation-based. Paper Fig. 10 ↗