船舶分段装配过程中，大型曲面薄板（如外板）放置在胎架上时，会受重力作用发生形变，将影响装配精度进而影响分段建造质量。为能预测给定胎架布局下大型曲面薄板的形变，本文建立了一种基于两阶段拉丁超立方采样和Transformer神经网络结构的代理模型（TSM-TLHS）。首先，设计了两阶段拉丁超立方采样，相较传统方法，能直接适用于形状不规则薄板的采样。同时，建立了包含多头注意力模块和位置编码的Transformer代理模型，综合考虑胎架位置与胎架布置点位移对薄板形变的影响。实际案例结果显示，本文提出的TSM-TLHS方法的预测误差仅为0.061mm，且满足现场装配对薄板形变的预测精度需求，便于船厂及时对分段进行反变形补偿，从而确保装配质量。

During the block assembly, when large curved thin plates (such as outer plates) are placed on the jigs, they undergo deformation due to the force of gravity This deformation affects the accuracy and quality of the block assembly in shipbuilding. In order to predict the deformation of these large curved thin plates within a given jig layout, this paper introduces a Transformer-based surrogate model with two-stage Latin hypercube sampling (TSM-TLHS). Primarily, compared to traditional approaches, the two-stage Latin hypercube sampling method enables direct sampling of irregularly shaped thin plates. Simultaneously, a Transformer-based surrogate model incorporating multi-head attention modules and positional encoding is used to comprehensively consider the impact of jig positions and corresponding node displacements on thin plate deformation. Real-case results demonstrate that the prediction error of this TSM-TLHS method is only 0.061mm, meeting the on-site assembly precision requirements for predicting plate deformation. This facilitates timely anti-deformation compensation by block in shipyards, ensuring assembly quality.