Data-Driven Ship Parametric Rolling Prediction: LightGBM Modeling and SHAP-Based Interpretability Analysis

Abstract

Within the framework of the International Maritime Organization (IMO) Second-Generation Intact Stability Criteria, parametric roll (PR) is recognized as an important dynamic stability failure mode that may occur when ships operate in waves. Efficient surrogate approximation of the PR response is therefore relevant to stability assessment under the loading and environmental conditions represented in the database. This study develops a LightGBM-based surrogate model for estimating the COMPASS-calculated parametric-roll response of a large container ship. A COMPASS-generated numerical dataset covers 36 retained loading conditions and includes five loading-condition parameters—displacement, LCG, VCG, RMI, and Cm—together with ship speed, SWH, and AZP. The regression target is the parametric-roll amplitude calculated by COMPASS for each parameter combination. The predictive performance of LightGBM is evaluated using the coefficient of determination and root mean square error and is compared with that of selected benchmark machine learning algorithms. On the held-out loading-condition test set, LightGBM achieves a higher coefficient of determination and a lower root mean square error than the comparison models, indicating favorable performance within the investigated condition space. SHapley Additive exPlanations (SHAP) are further employed to quantify the contribution of individual input features and to examine the relationships between the governing parameters and the predicted parametric roll response. The identified feature effects are generally consistent with the expected physical relationships associated with ship roll behavior. These results demonstrate the feasibility of using LightGBM as an interpretable surrogate for COMPASS-calculated parametric-roll responses within the investigated numerical dataset. Further validation using additional ship types, operating conditions, and experimental or full-scale data is required before the model can be applied more broadly in engineering practice.