This repository contains trained Physics-Informed Neural Network (PINN) models used in the comparative analysis of accuracy and computational efficiency for solving the two-dimensional Helmholtz equation. The models correspond to the network architectures analyzed in the associated publication, including variations in the number of hidden layers (1–3) and neurons per layer (25–75). All models were trained using a fixed random seed to ensure reproducibility of the reported accuracy metrics. The remaining hyperparameters were fixed according to the optimal configuration identified during the hyperparameter optimization stage: learning rate α=10−2, number of collocation points N=25, number of layers L=3, and sine activation function. The trained models enable reproducible evaluation of solution accuracy and inference-time performance, and can be directly used to reproduce the error metrics and computational benchmarks reported in the manuscript. Related publication Benchmarking Physics-Informed Neural Networks and Boundary Element Method: Accuracy–Efficiency Trade-offs in Wave Scattering