Learning highly oscillatory optical fields with Fourier feature networks.

Area of Science:

• Photonics and optical engineering
• Machine learning applications
• Computational physics

Background:

• Accurate modeling of physical perturbations in optical systems is crucial for photonic device design.
• Current characterization methods are often computationally intensive and time-consuming.
• Understanding how physical changes affect light transmission is key for advanced optical technologies.

Purpose of the Study:

• To introduce a data-efficient machine learning framework for modeling perturbation-dependent transmission matrices in optical systems.
• To overcome the spectral bias limitations of standard neural networks in capturing high-frequency phase changes.
• To create a continuous, differentiable digital twin of the optical system for robust characterization.

Main Methods:

• Developed a machine learning framework encoding perturbations into a Fourier Feature basis.
• Utilized a compact multi-layer perceptron for high-fidelity mapping from sparse training data.
• Employed experimental data from a mechanically deformed multimode fiber for model training and validation.

Main Results:

• Achieved a 0.996 complex correlation with experimental ground truth data.
• Improved phase accuracy by an order of magnitude compared to standard neural networks.
• Demonstrated superior performance with significantly fewer model parameters.

Conclusions:

• The proposed framework offers a computationally efficient and highly accurate method for characterizing complex optical media.
• The Fourier Feature encoding successfully addresses spectral bias, enabling precise phase change resolution.
• The 'digital twin' approach provides a robust tool for real-time monitoring and design in dynamic optical environments.