RecNet: advanced encoder-decoder architecture for SHG-FROG pulse reconstruction with enhanced noise immunity and...

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Signal Processing

Background:

• Frequency-resolved optical gating (FROG) is crucial for characterizing ultrashort laser pulses.
• Second Harmonic Generation (SHG) FROG is a common technique, but its trace reconstruction is sensitive to noise.
• Existing reconstruction algorithms often struggle with noisy data and lack interpretability.

Purpose of the Study:

• To introduce RecNet, a novel convolutional neural network for reconstructing SHG-FROG traces.
• To enhance reconstruction robustness and interpretability by incorporating domain knowledge constraints.
• To demonstrate RecNet's superior performance compared to existing methods.

Main Methods:

• Developed RecNet, an encoder-decoder convolutional neural network architecture.
• Implemented a domain knowledge-embedded loss function to enforce noiseless sample constraints.
• Utilized an architecture that matches trace dimensions with intermediate representations for constraint application.
• Conducted comparative studies against classical algorithms (PCGPA) and other neural networks.

Main Results:

• RecNet significantly improves reconstruction accuracy compared to PCGPA and non-constrained neural networks.
• The model demonstrates a higher convergence ratio in trace reconstruction.
• Experimental validation confirms RecNet's superior performance and robustness to noise.

Conclusions:

• RecNet offers a robust and accurate solution for SHG-FROG trace reconstruction.
• Incorporating domain knowledge into neural network loss functions is effective for optical signal processing.
• RecNet represents a significant advancement in ultrafast optical pulse characterization.