Generative models, surrogate optimization, and reinforcement learning are flipping the colloidal self-assembly workflow from forward to reverse.