Rogue hunters: A dual-branch deep learning framework with regional insight and gradient-enhanced loss for optical rogue wave prediction

This paper introduces a dual-branch neural network for real-time prediction of optical rogue waves, which are complex nonlinear phenomena in fiber-optic supercontinuum generation. Our framework employs a threshold-based segmentation strategy to partition optical field data into high-intensity and low-intensity regions, allowing for specialized processing of extreme events and background components. The high-intensity branch utilizes a network of depthwise separable convolutions and a bidirectional Long Short-Term Memory (LSTM) with an attention mechanism, while the low-intensity branch uses a simplified LSTM-attention architecture. Simulation results demonstrate that our model accurately reproduces key dynamical features, such as soliton formation and energy exchange. A comparative analysis shows that the dual-branch strategy achieves a temporal prediction error of 0.00641, reducing the error by approximately 70.7% compared to a conventional holistic model, while also decreasing training time by 16.7%. These results highlight the framework’s potential for real-time monitoring and design of high-power nonlinear optical systems.