Can AI automatically design new materials by working backward from desired properties?

Materials scientists are inverting the discovery process: instead of testing what existing compounds do, they're asking AI to propose new ones that meet specific targets. This review covers how generative models (VAEs, diffusion models, flows, autoregressive networks) learn crystal structure patterns from databases, then produce candidates constrained by physics and chemistry. It examines how combining diverse data modalities—structures, spectra, thermodynamics, text—builds richer representations, and how reinforcement learning and Bayesian optimization close the loop between prediction and synthesis. Key obstacle: models often propose mathematically valid but practically impossible structures; the review catalogs failure modes and evaluation standards to separate hype from real discoveries.