A Comparison of Deep Generative Models for Inverse Design of Steel Frames

Abstract

The design process of steel structures involves balancing multiple geometric and mechanical parameters to achieve structural efficiency and compliance with design standards. Traditional design method is an iterative forward tria I‐end‐error procedure that often relies on extensive numerical simulations, which are computationally expensive. In this work, we present and compare two deep generative models, conditional autoencoder (CAE) and variational conditional autoencoder (CVAE) for learning and exploring the design space of steel frames. The models are trained using a tabular dataset containing several examples calculated with a parametric model and Eurocode 3. The models enable the conditional generation of feasible frame geometries based on selected performance criteria, such as bending and shear capacity, while capturing the underlying distribution of optimal solutions. This approach provides a data‐driven methodology for accelerating the conceptual design process, reducing computational cost, and supporting engineers in identifying optimal frame configurations that meet structural requirements.