Title: Automating the Search: A Survey of Generative AI, Multimodal Integration, and Closed-Loop Systems in Inverse Materials Design

Inverse materials design is fundamentally transforming the field by moving away from forward prediction methods toward the targeted generation of candidate materials that meet specific objectives while adhering to physical constraints. This review examines recent progress in generative crystal structure modeling, multimodal learning techniques, and closed-loop design pipelines specifically for crystalline solids.

We analyze how contemporary generative models leverage chemical-structural priors extracted from extensive databases to facilitate the controllable sampling of periodic structures. The discussion compares several leading architectural classes, including variational autoencoders, normalizing flows, autoregressive models, and diffusion models. Special emphasis is placed on the mechanisms used to enforce feasibility constraints and physical priors throughout the design workflow. These include decisions regarding data representation, training objectives, guidance during the sampling phase, and subsequent screening and relaxation processes.

Furthermore, the article explores how multimodal learning integrates a wide array of materials data—such as crystal structures, thermodynamic and electronic properties, microscopy and spectroscopy images, processing contexts, and scientific literature—to build a more universal and transferable representation of chemical space.

The review also investigates various inverse-design strategies, with a focus on approaches that combine conditional generation with latent optimization, Bayesian optimization, reinforcement learning, and active learning. Finally, we address common failure modes encountered in these systems, such as surrogate exploitation, diversity collapse, distribution shift, and the gap between theoretical stability and actual synthesizability. To conclude, we propose evaluation standards for discovery-grade assessments, emphasizing staged reporting on validity, novelty, uniqueness, stability, and cost.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC