Reinforcement Learning-Based Optimization of Ku-Band Low-Noise Amplifier.

Area of Science:

• Microwave Engineering
• Machine Learning Applications
• Semiconductor Device Design

Background:

• Low-noise amplifiers (LNAs) are critical components in RF systems, requiring precise design for optimal performance.
• Traditional LNA design is complex and time-consuming, often relying on iterative manual adjustments.
• Reinforcement learning (RL) offers a promising approach for automating complex engineering design processes.

Purpose of the Study:

• To automate the design optimization of a wideband LNA operating in the Ku-band (12-18 GHz).
• To apply proximal policy optimization (PPO), a reinforcement learning algorithm, to microwave circuit design.
• To achieve specific performance targets including return loss (S11, S22) and noise figure (NF).

Main Methods:

• Utilized a two-stage LNA architecture with transmission lines (TLs) and capacitors as design parameters.
• Employed proximal policy optimization (PPO) for automated design optimization, adjusting 29 parameters.
• Set target specifications: S11/S22 < -10 dB across 12-18 GHz, NF < 2 dB.
• Performed 20,140 simulations for training using a 150 nm GaN HEMT process model.

Main Results:

• The RL training process showed appropriate convergence of reward and loss functions.
• The automated design successfully met the target specifications for S11, S22, and NF.
• The optimization process demonstrated the efficacy of PPO in complex microwave circuit design.

Conclusions:

• Reinforcement learning, specifically PPO, is an effective method for automated design optimization of wideband LNAs.
• The study successfully demonstrated the feasibility of achieving stringent RF performance targets through automated design.
• This approach has the potential to significantly reduce design time and effort in microwave engineering.