Abstract
This study proposes a surrogate model-based two-stage optimization method for the structural design of the pump tower installed within the LNG (Liquefied Natural Gas) cargo tank. The optimal design of the pump tower must simultaneously satisfy multiple classification-rule constraints in a mixed-variable design space that combines tripod cross-section geometry variables with strut layout variables, and it inherently requires repeated, costly high-fidelity structural analyses to evaluate structural responses for each design alternative. To address this limitation, this study integrates the following four elements into a single consistent design method: (1) a parametric representation of the pump tower geometry; (2) automated generation of structural-analysis input files reflecting sloshing and thermal load conditions; (3) the development of a probabilistic surrogate model that quantitatively evaluates compliance probability by modeling the distribution of member-wise stress ratios; and (4) an NSGA-II-based two-stage optimization procedure that aligns with the hierarchical structure of the design variables. The probabilistic surrogate model in this study consists of a classification-based surrogate model that directly learns the compliance boundary, a distribution-parameter regression surrogate model that predicts the log-normal distribution parameters of member-wise stress ratios, and an ensemble surrogate model that combines the two via soft voting. As a result of the optimization, the first-stage optimization of the strut layout and member configuration yielded a 17.012-ton first-stage optimized design alternative, representing a 13.257% structural weight reduction relative to the manual design (19.612 tons); the second-stage optimization precisely adjusted the tripod cross-section geometry variables and converged to a 16.486-ton final optimized design alternative, achieving a cumulative structural weight reduction of 15.939% relative to the manual design. In the high-fidelity structural analysis verification of the stage-wise optimized design alternatives, the member-wise stress ratios were confirmed to fall sufficiently below the allowable limit of 1.0 across all eight LR (Lloyd’s Register) rule evaluation items. These results quantitatively demonstrate that the proposed method is a practical design-support method that can consistently secure both feasibility according to the probabilistic surrogate model and structural safety in high-fidelity structural analysis.
