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

This paper presents RecNet, an innovative convolutional neural network designed for reconstructing cecond harmonic generation frequency-resolved optical gating (SHG-FROG) traces. Unlike conventional approaches, RecNet incorporates noiseless sample constraints through a domain knowledge embedded loss function, enhancing the network's robustness to noise and interpretability. The encoder-decoder architecture is intentionally selected to match the dimensions of the trace diagram with intermediate representations, facilitating the effective application of these constraints. Comparative studies show that RecNet surpasses classical algorithms like PCGPA and network models that do not incorporate domain knowledge constraints in reconstruction accuracy and convergence ratio. Experimental results further confirm the superiority of RecNet.