Abstract: To achieve high-accuracy and high-efficiency solutions for acoustic scattering from underwater elastic targets, a physics-informed neural network–based acoustic scattering modeling method is proposed, with emphasis on improving convergence performance through the network architecture, activation functions, and loss weight distribution. The influence of the number of configuration points in each region on prediction performance is analyzed. The physical constraints of the acoustic pressure field in the fluid domain and the displacement field in the elastic body domain are decoupled to construct parallel and sequential network architectures. Gain adjustment is applied to output values to balance the characteristic dimension differences of different physical fields. An adaptive weight method and Snake activation function are introduced to further enhance model efficiency. Numerical simulations are conducted on two-dimensional elastic targets. The results show that when the target is circular and the frequency is 2409 Hz, the convergence efficiency of the decoupled parallel model is 78.8% higher than that of the traditional model. As the frequency and shape complexity increase, the decoupled parallel model demonstrates significant improvements in convergence efficiency and generalization ability.