<p>The soil arching effect under unloading conditions, such as trench excavations, tunnels, earth dams, buried pipelines and pile-supported embankments, was investigated using trapdoor testing data in sand. A structured database was established, including 5 soil properties (median particle size, coefficient of uniformity, relative density, void ratio, and internal friction angle), 2 structural parameters (trapdoor width and fill height), and key points of the ground reaction curve, including minimum and ultimate soil arching ratios with the corresponding normalized trapdoor displacements. Convolutional neural network models were developed to predict these parameters and reconstruct the full ground reaction curve. Hyperparameters were optimized using Bayesian optimization and particle swarm optimization, resulting in four predictive models. The model performance was evaluated through comparative validation, monotonicity analysis, sensitivity assessment, and robustness evaluation. The Bayesian optimization-based model achieved the highest predictive accuracy, maintained physical consistency, and showed stable performance under varying input conditions. The optimized model was implemented as a machine learning module to support single-case and batch predictions. Rapid visualization of ground reaction curves was achieved, supporting efficient data-driven analysis. The proposed framework combined a curated database, neural network-based curve prediction, and hyperparameter optimization, providing a robust approach for evaluating the soil arching effect.</p>