Transformer-Based Generative Model for Compliance Optimization of Planar Trusses

This study proposes a Transformer-based generative model for minimizing the compliance of planar trusses. Features of geometry and boundary-conditions are encoded as variable-length inputs, and design variables are produced through sensitivity-based learning without labeled data. Compliance and volume sensitivities are incorporated directly into training, enabling stable convergence across diverse structural configurations. Numerical experiments show that the proposed approach yields approximate optimal designs comparable to gradient-descent solutions, while requiring substantially lower computational cost. The model also generalizes to untrained loading and support conditions.