Topology optimization using machine learning

Abstract

This research presents an advanced methodology for structural design through the integration of topology optimization (TO) and machine learning (ML), addressing the limitations of traditional design processes in terms of efficiency and scalability. While the initial approach using TOPY for dataset generation proved computationally inefficient, this work focuses on overcoming data challenges by identifying and leveraging alternative tools for 3D dataset creation. Specifically, the DL4TO library was utilized as a starting point for dataset development, although its limited dataset size necessitated further data augmentation. Finite element methods were explored, including PETSc (found computationally inefficient), COMSOL (limited by data export issues), and Altair, which demonstrated robust capabilities for simulation and data export. The study integrates the data from DL4TO with additional 3D datasets generated using Altair to train a deeper 3D U-Net architecture, optimizing material distributions under varied load and boundary conditions. Preliminary results show that this enhanced architecture significantly improves prediction accuracy and scalability for 3D structural optimization tasks. The proposed methodology not only streamlines the optimization process but also establishes a framework for integrating high quality datasets and ML-driven TO models, paving the way for future advancements in engineering design automation.