Variable-Resolution Topology Optimization Based on Dynamic Vision Transformer

The integration of a Dynamic Vision Transformer (ViT) into topology optimization (TO) enables adaptive resolution adjustment, addressing the computational inefficiency inherent in fixed-resolution optimization methods. This paper proposes a Dynamic ViT-based Variable-Resolution Topology Optimization (DT-VRTO) framework, which leverages an adaptive network architecture to implement stepwise optimization from low to high resolution. The core contribution of this work lies not in designing brand-new neural network architectures, but in integrating super-resolution-assisted resolution mapping and dynamically configured ViT surrogate modeling into a unified variable-resolution TO framework. By coupling low-resolution optimization, super-resolution-guided high-resolution initialization, and online surrogate updating, the proposed DT-VRTO achieves a favorable balance between computational efficiency and optimization accuracy across multiple resolution levels. The method is validated on both 2D and 3D compliance minimization benchmarks to verify its computational efficiency and cross-case transferability. Numerical results demonstrate that DT-VRTO substantially reduces the computational cost compared with conventional high-resolution TO. In the 2D benchmark, the proposed method reduces the iteration count by 53.1% and structural compliance by 11.8%. For the 3D case, it decreases iterations by more than 19% while maintaining a slight improved compliance performance.