eCNNTO: Accelerating Topology Optimization with a Generalizable ConvNet

Topology Optimization (TO) is a powerful design method, but its efficiency is often hampered by the large number of iterations required, each involving computationally intensive finite element analysis. This bottleneck becomes particularly severe when high-resolution designs necessitate dense meshes. To tackle this, a new method called eCNNTO (element-based Convolutional Neural Network for Topology Optimization) has been proposed. eCNNTO builds upon prior work that used Deep Belief Networks to predict near-optimal densities for individual elements, thereby skipping many iterations. However, that approach lacked consideration for spatial correlations between neighboring elements, potentially leading to disconnected features in the final designs. eCNNTO addresses this by employing a Convolutional Neural Network (CNN) with residual connections, which inherently captures these spatial relationships. Furthermore, eCNNTO introduces a novel training strategy. Instead of using early density histories, the training dataset consists of final-stage density histories, which not only improves optimization efficiency but also reduces the required training data size. The method demonstrates strong generalization capabilities, performing well across problems with varying boundary conditions, loading cases, design domain geometries, mesh resolutions, and non-design domains. It achieves significant iteration reductions, up to 90% in 2D and 97% in 3D examples, showcasing its potential to greatly accelerate TO procedures.