Deep-learning-enabled development of a high-dynamic-performance CFRP damping boring bar for chatter suppression

In the development of composite boring tools, the structural design of the boring bar plays a decisive role in tool dynamic performance. Composite damping boring bars involve strong multi-material coupling and pronounced nonlinear responses; consequently, conventional design approaches often rely on extensive experiments and repeated parameter tuning, resulting in long development cycles and high costs. To address this issue, this paper proposes a two-stage multi-fidelity LSTM deep learning network (TS-MF-LSTM) to predict the dynamic responses of CFRP damping boring bars. The proposed method adopts a two-stage learning strategy: it is first pre-trained on a large-scale simulation dataset and then fine-tuned via a residual-correction network to improve high-fidelity predictions. For natural frequency prediction, the model achieves R 2 = 0.9624 and MAE = 18.49 Hz; for dynamic stiffness prediction, it achieves R 2 = 0.9836 and MAE = 0.0402. After model training, particle swarm optimization (PSO) is further integrated to maximize the dynamic stiffness and natural frequency of the boring tool. By leveraging TS-MF-LSTM predictions, the carbon-fiber layup angle and layer-thickness ratio are rapidly optimized, enabling efficient structural parameter design. To validate the proposed approach, two boring bars were manufactured based on the optimization results: one corresponding to the optimizer-derived optimal solution and the other representing a suboptimal design with a predicted dynamic performance approximately 10% lower. Dynamic tests under identical machining conditions show that the optimal design exhibits superior vibration behavior in experiments, in close agreement with the model predictions.