Multi-Cluster Physics-Information Fusion Network for Vessel Drag Prediction

Objectives To address the high-precision drag prediction requirements for diverse ship types during the design phase under data feature missing scenarios, this study proposes a prediction method integrating missing-data-robust feature selection and a multi-cluster physics-informed expert network, aiming to enhance the model’s adaptability to ship-type heterogeneity and incomplete data. Methods Based on 6,082 bare-hull ship model towing tank test cases covering 17 ship types, we employed a missing-data-robust random forest regressor combined with partial correlation coefficients to select 7 independent dimensionless ship parameters. Missing values were addressed via decision-tree-based tiered imputation. A weighted K-Means clustering algorithm partitioned the feature space into 5 sub-regions, with each cluster independently trained using physics-informed local expert networks. The loss function incorporated class-balanced weighting and physics-informed regularization terms. A multi-seed ensemble method reduced prediction variance, while friction resistance was calculated using the Prandtl-Schlichting formula and International Towing Tank Conference (ITTC) roughness correction. Results On the test set, the R² averaged 0.93 (peak 0.987). For benchmark models KVLCC2, KCS, and DTMB5415, the Mean Absolute Percentage Error (MAPE) across all speeds was 8.2%, 6.0%, and 2.2%, respectively. For non-bare-hull special-purpose vessels, the full-speed MAPE remained below 10%.Conclusions This study integrates missing data robustness, feature interaction modeling, adaptive clustering, and physics-constrained embedding, achieving cross-ship-type prediction accuracy surpassing traditional classification-based models. The developed surrogate model balances prediction precision and physical reliability, providing an efficient tool for preliminary ship design with significant engineering value.