A novel deep learning approach for accurate and efficient design of LNOI power splitters

Abstract Photonic Integrated Circuits (PICs), owing to their high speed, low power consumption, and compact structure, lie at the core of modern optoelectronic technologies. The design of these circuits requires high accuracy and intensive computational cost. In this study, a novel Deep Neural Network (DNN)-based framework is proposed for designing and predicting the performance of arbitrary-ratio power splitters on the Lithium Niobate on Insulator (LNOI) platform. A dataset constructed using fundamental geometric parameters such as width, height, length, and auxiliary dimensions was processed with the proposed DNN model, yielding high prediction accuracy. The model achieved strong agreement in the training, validation, and testing stages, with R² values of 0.95, 0.97, and 0.97, respectively. The corresponding error metrics were RMSE = 3.08, 2.4, and 2.5, and MAPE = 4.02%, 3%, and 3.1%, respectively. Extensive analyses across various epoch numbers (500–10,000), batch sizes (2–64), and optimizers (Adam, SGD, RMSProp) revealed that the Adam optimizer, with 5,000 epochs and a batch size of 64, achieved the optimal balance between accuracy, convergence speed, and generalization. Furthermore, a detailed analysis of the influence of input parameters on outputs revealed that L 1 and W were the most critical factors. The trained model was also validated on an independent dataset from the literature, demonstrating excellent generalization ability with R = 0.991, RMSE = 1.98, and MAPE = 3.42%. To facilitate practical use of the proposed framework, an interactive MATLAB application was developed, enabling both forward prediction of power-splitting ratios from user-defined geometric inputs and inverse design of optimal parameters corresponding to a target output ratio through an integrated DNN–optimization workflow. This tool significantly accelerates device evaluation and design-space exploration, making the methodology readily applicable to real-world photonic design tasks. These results indicate that the proposed approach not only accelerates the design process but also enhances the understanding of input-output relationships, thereby providing a reliable methodology for photonic device optimization.