Applications and Improvements of Multi-Scale Graph Neural Networks in Modeling Spatial Distribution of Optoelectronic Signals

Given the fast-growing needs for optoelectronic technology in numerous high-end applications (such as high-precision communication, optical sensing and imaging, etc.), accurately characterizing and predicting the spatial distribution of photoelectric signals has become the major challenge for scientific research and industrial applications. Focusing on the multi-level graph neural network being used, this study separately proposes cross-scale information fusion, attention gating and topological embedding modules, extracts the feature distribution of photoelectric signals at different spatial scales through multi-scale graph convolution, and adopts a learnable adjacency and dynamic weighting strategies to mine local and global associations. Within this, the residual attention mechanism is adopted to realize both interactive expression enhancement and gradient transmission stability for key node feature vector, and with investigations such as self-supervised pre-training, multi-loss collaborative optimization, the network achieves a balance between training efficiency and generalization performance for photoelectric signals modeling at multiple levels and scales. Multi-scale graph neural network is proposed based on the dataset of simulated scenarios and the experimental results indicate that it has a significant improvement on prediction accuracy, robustness and generalization ability of spatial distribution of photoelectric signals.