Generative Design of Additively Manufactured Steel Dampers Using a Feasible Geometry-Constrained VAE–GAN

Abstract Generative design based on deep learning has enormous potential for developing innovative structural components, such as steel dampers for seismic design. Nevertheless, an important challenge to real-world implementation is that current models frequently produce disconnected structures and physically infeasible designs that cannot be manufactured. This study overcomes this fundamental challenge by introducing a deep generative framework that integrates a connectivity constraint. The manufacturability of generated lattice structures is directly implemented through a hybrid Variational Autoencoder–Generative Adversarial Network (VAE–GAN) model that includes a connectivity loss function. Multi-objective optimization produced a physically meaningful Pareto dataset that was used to train the model. The seismic performance of the generated designs is then rapidly predicted by a convolutional neural network, significantly reducing evaluation costs and providing a suitable alternative by comparing the performance of generated shapes. This framework is used to design high-performance steel dampers for seismic design. The optimal design is fabricated via metal additive manufacturing (AM) and tested in real experiments and via nonlinear finite element analysis, demonstrating close agreement with predictions and a robust efficiency gain over conventional dampers, with parity with honeycomb dampers on a per-volume basis. The experimental campaign further demonstrated a progressive, multi-path energy-dissipation mode enabled by the lattice’s structural redundancy, which is generally challenging to achieve with conventional standardized dampers. This method extends the realistic design space beyond parametric designs created individually, reduces assessment costs, and incorporates manufacturing feasibility into the generation process. By ensuring that manufacturing feasibility is a crucial component of the generative process, this research offers a comprehensive, validated workflow that closes the gap between generative design and practical engineering.