MACHINE LEARNING-ASSISTED TOPOLOGY OPTIMIZATION OF LIGHTWEIGHT MECHANICAL STRUCTURES FOR ADDITIVE MANUFACTURING

Abstract

The increasing demand for structural lightweighting in aerospace, automotive, and biomedical sectors has driven significant advances in topology optimization (TO) and additive manufacturing (AM). This review examines the integration of machine learning (ML) into topology optimization workflows, addressing the persistent computational bottleneck of iterative finite element analysis (FEA) that limits conventional TO. A comprehensive taxonomy of traditional TO formulations including element-based continuous methods (SIMP, RAMP), discrete evolutionary approaches (ESO, BESO), and boundary/interface tracking techniques (Level Set Method) is presented, highlighting their numerical instabilities such as checkerboard patterns, mesh dependency, and gray transition zones. The review systematically analyzes three primary machine learning paradigms: forward surrogate networks, operator learning architectures such as U-Net Fourier Neural Operators (FNOs), and Physics-Informed Neural Networks (PINNs). Comparative analysis reveals that ML-assisted TO achieves computational speedups ranging from 2× to 100× while maintaining structural fidelity and compliance accuracy. Furthermore, the integration of additive manufacturing process constraints including overhang angle limitations, self-supporting requirements, thermal distortion, and build direction anisotropy directly into neural network loss functions enables the generation of print-ready geometries without sacrificial supports. Industrial case studies demonstrate mass reductions of 60–77% with validated structural integrity, while stress concentrations decrease by up to 50%. Despite challenges in generalization, training time, and non-convex loss convergence, ML-assisted TO offers a transformative pathway toward sustainable, manufacturable lightweight structures.