These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Topology optimization (TO) is a critical technique in engineering design for creating lightweight structures with optimal stiffness and strength. However, the high computational cost of traditional TO methods limits their use in early design stages where rapid iteration and flexibility are essential. Recent advances in machine learning (ML), particularly deep learning, offer potential solutions to these challenges. In particular, the integration of generative models with TO can efficiently produce high-fidelity designs. In this study, we adapt a generative ML framework from 3D computer vision to structural design applications. Our approach uses an implicit shape representation conditioned by a code vector, allowing a single network to represent a wide range of shapes and interpolate smoothly between them. The primary goal is to develop a generative auto-decoder network that uses model boundary conditions as input parameters to generate new, accurate design proposals. To construct the training dataset, we established a parameterized finite element analysis (FEA) model, including a unit force with variable position and orientation as the load, along with adjustable support points. A parameterized volume constraint accounts for different load levels. Random samples were generated using the Latin Hypercube method to ensure uniform coverage of the parameter space. The TO was performed with a combined objective function aiming for both high stiffness and low stress concentration. Model evaluation involved comparing generated shapes with unseen test cases, focusing on accuracy and stress performance through simulation. The results show that the trained generative model effectively produces structurally optimized designs with high accuracy and low memory requirements. This capability makes it suitable for the rapid generation and validation of early-stage design concepts for lightweight components. In addition, the model's ability to adapt to changing boundary conditions in near real-time indicates its potential for applications requiring rapid design iteration, such as instantaneous TO and conceptual design workflows. By integrating generative models with TO, this work provides a practical approach to reducing computational costs and increasing design flexibility in the early stages of engineering design. This is in line with current advances in ML-based TO methods and contributes to improving the efficiency of engineering design practices.