Due to samples scarcity, high cost, and deformation complexity, structural optimization of collision problem was a challenging role. Some new technologies were introduced to solve the problem of sample scarcity, such as unsupervised learning, optimal Latin design, and dynamic surrogate model. Inspired from the semantic vector in large language models, new method based on mechanics experience was proposed, to facilitate structural samples mining. Compared to conventional optimization analysis, newly proposed methodologies focused on enhancing information efficiency and minimizing computational workload, enabling rapid and high-quality solutions for strongly nonlinear problems. Two methodologies with different priority targets and procedures were applied, to optimize the simplified ship-side structure that subjected to external bow-tip collision. The first methodology prioritizes solution speed, integrated optimal Latin design, and unsupervised learning to mining samples. Specifically, by utilizing collision-related empirical parameters as classification vector to enhance samples mining effectiveness, surrogate models were trained using representative samples and computational burden was reduced significantly. The second methodology devoted to achieve depth and speed balanced optimization. It implemented factorial design and unsupervised learning method to select the initial representative samples for model training. Based on initial surrogate model, superior region and the direction of gradient descent were searched to generate the new samples and update the surrogate model. Finally, characteristics of the two methodologies were compared. With the multi-objective optimization, specification verification, unsupervised learning effectiveness assessment, and comparison with Bayesian Optimization, the rigor and feasibility of the newly proposed optimization methodology were tested.
Qiu et al. (Sun,) studied this question.