Abstract: Redistribution Layer (RDL) technology plays a critical role in 2.5D and 3D packaging, serving as a key component for inter-chip connections. Traditional design methods often face challenges of low computational accuracy or long simulation times when dealing with complex multilayer structures. To address these issues, this paper proposes an innovative RDL transmission line design optimization method combining a Multilayer Perceptron (MLP) neural network with Bayesian optimization. Firstly, a dataset containing various design and parasitic parameters is generated through simulations, and the CLnet neural network model is trained to accurately predict parasitic parameters of different RDL transmission line types. Then, Bayesian optimization is used to reverse-design transmission line parameters that meet specific impedance requirements, efficiently searching for the optimal design. Experimental results show that with two degrees of freedom in design parameters, this method is over twice as efficient as electromagnetic simulations and provides multiple transmission line designs that meet performance requirements. It is fast, accurate, and adaptable to various substrate structures, showing broad application potential.