eCNNTO: A Highly Generalizable ConvNet for Accelerating Topology Optimization

Topology optimization (TO) is essential for high-performance structural design but is computationally expensive due to the need for repeated finite element analysis (FEA) across many iterations. Existing neural network-based acceleration methods often struggle with either poor structural connectivity (due to ignoring spatial correlations) or a lack of generalizability to new boundary conditions and geometries. This paper introduces eCNNTO to address these bottlenecks by leveraging spatial information through convolutional neural networks (CNNs).

eCNNTO builds upon the concept of learning element-wise density evolution. Unlike previous approaches that treat elements independently, eCNNTO uses a CNN architecture with residual connections to explicitly model spatial correlations among neighboring elements. The model takes a small window of element density history (from the early stages of SIMP) as input and predicts the near-optimal final density. A key innovation is the training strategy: the model is trained on density histories from the final stages of optimization rather than early ones, which improves efficiency and reduces the volume of training data required.

The authors demonstrate that eCNNTO significantly accelerates the optimization process, achieving up to 90% reduction in iterations for 2D problems and up to 97% for 3D problems. By incorporating spatial context, the model effectively suppresses common defects like isolated structural pieces and checkerboard patterns. Furthermore, the model exhibits strong generalization capabilities, successfully applying to problems with different boundary conditions, loading cases, and mesh resolutions without needing retraining.