Addressing data scarcity in optical matrix multiplier modeling using transfer learning.

Area of Science:

• Photonics
• Optical Computing
• Machine Learning

Background:

• Training neural network (NN) models for optical matrix multipliers often requires extensive experimental data.
• Data scarcity poses a significant challenge in developing accurate photonic devices.

Purpose of the Study:

• To evaluate transfer learning for training NN models for Mach-Zehnder interferometer mesh-based optical matrix multipliers.
• To address experimental data scarcity in NN model training for photonic applications.

Main Methods:

• Pretraining NN models using synthetic data from an analytical model.
• Fine-tuning the pretrained models with limited experimental data.
• Employing regularization techniques and ensemble averaging.

Main Results:

• Transfer learning significantly reduced modeling errors compared to standalone analytical or NN models.
• Achieved <1 dB root-mean-square error on 3x3 matrix weights.
• Successfully trained models using only 25% of available experimental data.

Conclusions:

• Transfer learning is an effective strategy to overcome data scarcity in training NN models for optical matrix multipliers.
• The proposed method enables accurate photonic device modeling with reduced experimental effort.
• This approach enhances the practicality of developing complex photonic integrated circuits.