Conventional multi-core fiber design methodologies face significant challenges, including inefficient processes and high computational requirements, particularly when addressing complex multi-objective optimization problems that depend on parameter sweeps and trial-and-error iterations. The non-uniqueness problem arising from multi-parameter coupling further compounds these limitations. This research presents a machine learning (ML)-based collaborative design framework to address these challenges for randomly coupled multi-core optical fiber (RC-MCF) and few-mode MCF (FM-MCF), aiming to enhance design efficiency, minimize computational resource utilization, and resolve the complexities associated with multi-parameter interactions in fiber design.

For RC-MCFs, a forward prediction model is constructed by integrating neural networks (NNs) and optimization algorithms, comprising three sequential stages: first, a classification model (NN 1) is trained on structural parameters including core radius (r), refractive index difference (Δ), and twisting peak rate (T_p) to predict the optimal core spacing (Λ) with high accuracy; second, a sparrow search algorithm-optimized random forest model (SSA-RF) is employed to classify spatial mode dispersion (SMD) into ordered labels, thereby enhancing feature representation for subsequent analysis; third, a particle swarm optimization-optimized random forest model (PSO-RF) utilizes the enhanced features—including the predicted Λ and SMD labels—to regress SMD values, capturing the complex nonlinear relationships between structural and optical properties. For FM-MCFs, two inverse design strategies are proposed: the first is an independent neural network model (NN 2), a five-layer NN with L2 regularization that directly maps target optical performance parameters—such as inter-core crosstalk(IC-XT), differential mode group delay (DMGD), effective area (A_eff), and bending loss (BL)—to structural parameters including r₁, r₂, w₁, Δ₁, α, Λ, and d_OCT; the second is a NN-PSO joint model (NN 3+PSO), where a forward NN (NN 3) first learns the mapping from structural parameters to performance metrics, and this model is then integrated with a particle swarm optimization (PSO) algorithm to serve as a fitness function for searching optimal structural parameters that meet predefined target performance criteria.

The implementation of these frameworks demonstrates significant effectiveness: for RC-MCFs, the NN 1 model achieves a core spacing prediction accuracy of approximately 0.989, as validated by a confusion matrix, while the SMD prediction exhibits a Pearson correlation coefficient (PCC) of approximately 0.9561 [Fig. 5(b)], indicating strong alignment between predicted and actual values; notably, the ML model substantially improves computational efficiency, reducing prediction time from hours for traditional numerical simulations to milliseconds after training (Table 1), enabling rapid design iterations. For FM-MCFs, the NN 2 model achieves PCC values of 0.9811 for κ, 0.9974 for DMGD, and 0.9870 for A_eff, with a mean square error (MSE) of approximately 5×10^-4, while the NN 3+PSO model further enhances accuracy, achieving PCC values of 0.9988 for κ, 0.9996 for DMGD, and 0.9914 for A_eff, with an MSE of approximately 3×10^-5 [Fig. 10(a)]; significantly, all predicted designs satisfy the stringent BL requirement of less than 0.001 dB/km under standard bending conditions (R=140 mm), as verified by COMSOL simulations [Fig. 10(b)]. The ML frameworks collectively reduce computational resource consumption by over 90% compared to traditional methods, effectively resolving the non-uniqueness problem in multi-parameter optimization and demonstrating the effectiveness of feature enhancement and deep learning in modeling complex nonlinear relationships between fiber structures and optical performances.

This research demonstrates the successful implementation of ML-assisted design frameworks for RC-MCFs and FM-MCFs, achieving significant improvements in both design efficiency and prediction accuracy. The forward prediction model for RC-MCFs and the inverse design strategies for FM-MCFs illustrate the transformative potential of machine learning in addressing the limitations of conventional design approaches, offering a reliable and scalable solution for developing next-generation large-capacity optical transmission systems. Although the current models demonstrate robust performance across typical parameter ranges, future research will focus on enhancing prediction accuracy in extreme parameter regions through data augmentation techniques and model optimization—such as implementing weighted loss functions—and expanding the applicability of these frameworks to other complex fiber designs, thereby advancing the field of optical fiber engineering.