<p>Machine learning (ML) has emerged as a powerful paradigm for the design and optimization of lightweight structures, offering substantial advantages over traditional simulation-driven or rule-based approaches. This review synthesizes recent progress in ML methodologies, including supervised learning, unsupervised and generative modeling, and reinforcement learning (RL), and evaluates their roles in predicting structural responses, generating candidate geometries, and navigating constrained multi-objective design spaces. Central to these methods is the representation of geometry, and we examine point-based, vector-based, and image- or voxel-based encodings that enable efficient learning across diverse structural forms such as lattices, cellular materials, and architected metamaterials. We further compare optimization strategies that integrate ML surrogates, generative models, or RL policies with external search algorithms, highlighting how these approaches improve computational efficiency and expand the accessible design space beyond conventional topology optimization. The review also outlines emerging trends in physics-informed learning, latent-space optimization, and diffusion-based generation, which provide controllable synthesis of structures under prescribed mechanical or manufacturing constraints. Finally, we discuss experimental validation of ML-optimized designs fabricated via additive manufacturing, emphasizing the importance of model calibration, field-level comparisons, and manufacturability-aware design loops for bridging digital predictions and physical performance. Collectively, this review provides a unified framework for understanding ML-driven lightweight design and offers insight into future opportunities for scalable, data-efficient, and experimentally grounded structural optimization.</p>