<p>Machine learning (ML) has strong potential for soil settlement prediction, but determining hyperparameters for ML models is often intricate and laborious. Therefore, we apply Bayesian optimization to determine the optimal hyperparameter combinations, enhancing the effectiveness of ML models for soil parameter inversion. The ML models are trained using numerical simulation data generated with the modified Cam-Clay (MCC) model in ABAQUS software, and their performance is evaluated using ground settlement monitoring data from an airport runway. Five optimized ML models—decision tree (DT), random forest (RF), support vector regression (SVR), deep neural network (DNN), and one-dimensional convolutional neural network (1D-CNN)—are compared in terms of their accuracy for soil parameter inversion and settlement prediction. The results indicate that Bayesian optimization efficiently utilizes prior knowledge to identify the optimal hyperparameters, significantly improving model performance. Among the evaluated models, the 1D-CNN achieves the highest accuracy in soil parameter inversion, generating settlement predictions that closely match real monitoring data. These findings demonstrate the effectiveness of the proposed approach for soil parameter inversion and settlement prediction, and reveal how Bayesian optimization can refine the model selection process.</p>