CMTCN-MPF: A dual-stream deep learning model for real-time nonlinear vibration prediction of smart composite panels

Smart composite panels embedded with macro-fiber composites (MFCs) exhibit tremendous potential for active vibration control under large deformations. However, the prohibitive computational cost associated with nonlinear solution procedures severely constrains the design and debugging of corresponding control algorithms. To address this limitation, a CMTCN-MPF deep learning framework is proposed. A Condition-Modulated Temporal Convolutional Network (CMTCN) dual-stream architecture is constructed by introducing a dedicated condition modulation mechanism, which significantly enhances the interpolation capability within continuous parameter spaces. Subsequently, a topology-aware multidimensional physics-fusion (MPF) loss function is designed to improve the reconstruction accuracy of high-order spectral features and transient dissipation boundaries of nonlinear systems. Additionally, a cosine annealing learning strategy incorporating a warmup mechanism is introduced to prevent the complex high-dimensional optimization process from falling into local minima. Systematic accuracy evaluations, parametric space tests, ablation studies, and experimental comparisons are conducted on the proposed framework. The results demonstrate that the proposed CMTCN-MPF framework accurately captures the geometric nonlinear features of the structure under large deformation. Meanwhile, the single inference time for long-duration nonlinear responses is compressed to 5.66 ms, thereby providing a novel technical approach for the nonlinear vibration prediction and dynamic evaluation of related structures.