Objective Current ship hull design practices remain heavily reliant on designers’ expertise, parent ship data, and foreign commercial CAD software, leading to extended design cycles and limited capacity for innovation. This study aims to investigate the application of three-dimensional generative large-scale models in ship hull conceptual design by establishing an intelligent design framework to overcome the limitations of conventional design paradigms.

Method First, four mainstream three-dimensional generative large-scale models (Hunyuan 3D, Meshy, Rodin, and Tripo) were selected to generate hull samples through dual-modal inputs consisting of text prompts and contour sketches. A systematic evaluation was conducted to compare the accuracy of feature representation and geometric integrity under varying input conditions. Second, a quantitative surface quality assessment method based on slope detection was proposed. This method identifies abnormal curvature regions on hull surfaces by calculating half-breadth values at mesh element centroids and incorporating multiple constraint conditions. Two quantitative indices, the Special Odd Area ratio (SOA) and the Normal Odd Area ratio (NOA), were defined to facilitate an objective evaluation of the fairness of generated hull surfaces. Third, a combined smoothing optimization pipeline integrating Laplacian and Taubin algorithms was developed. The Laplacian algorithm mitigates local noise and stepped artifacts by repositioning mesh vertices toward the geometric centers of their neighborhoods, while the Taubin algorithm effectively suppresses volume shrinkage through the alternating application of positive and negative smoothing factors. Together, these algorithms synergistically enhance surface quality to satisfy the prerequisites for CFD analysis. Finally, hydrostatic calculations were performed to verify the engineering feasibility of the geometric models, followed by numerical simulations of resistance performance using STAR-CCM+.

Results A systematic analysis of resistance coefficient variations across three candidate hull forms within the Froude number range of Fr=0.20～0.30 identified an optimal hull form, C2, generated by Rodin, which exhibited the lowest total resistance coefficients under all operating conditions. CFD results demonstrate that the AI-generated hull forms, after post-processing optimization, meet both mesh quality and hydrodynamic analysis requirements, thereby preliminarily validating the technical feasibility of generative large-scale models in ship hull design.

Conclusion Three-dimensional generative large-scale models show considerable potential for application in ship hull conceptual design. The proposed quantitative surface quality evaluation method and hybrid smoothing algorithms provide essential technical support for the engineering application of AI-generated geometries. The research outcomes effectively facilitate hydrodynamic performance analyses, such as resistance evaluation, laying a foundation for the development of intelligent hull optimization design frameworks. This approach reduces reliance on parent ship data and designer experience while promoting the development of innovative hull forms.