AI驱动的材料发现 : 生成式模型与晶体逆向设计进展

晶体材料的逆向设计一直是材料科学面临的重大挑战，近年来生成式人工智能技术的发展为解决这一难题带来了新的契机。本文围绕晶体表示方法与生成模型架构2大核心问题，系统综述了该领域的最新研究进展。在晶体表示方法方面，重点评述了文本符号化表示、几何图周期性编码以及像素的实验成像与计算像素化等方法，深入分析了这些方法在可逆性、不变性和对称性编码等关键技术特性上的优劣。在生成模型架构方面，按照技术演进路线，详细介绍了基于潜空间的变分自编码器和生成对抗网络、扩散或流匹配的联合生成模型，以及Transformer自回归模型的工作原理与应用现状。最后，针对当前面临的表示方法优化、缺陷材料建模和评估标准统一等关键科学问题，本文展望了面向工业应用的未来发展方向，旨在为晶体材料智能设计提供系统性的理论参考。

The inverse design of crystalline materials aims to generate novel structures with targeted properties. However, achieving this objective remains a long-standing challenge in materials science. Unlike molecular systems that benefit from mature representations such as SMILES, crystalline materials impose stringent physical constraints, requiring representations that are simultaneously periodicity-aware, invertible, and invariant under symmetry operations. Emerging generative artificial intelligence (AI) technologies present transformative opportunities to address these complexities.

This mini-review systematically examines the latest advancements in this field, centering on two critical pillars: crystal representation methods and generative model architectures. In terms of crystal representations, we critically evaluate text-based symbolic encodings (e.g., SLICES), geometry-based periodicity-aware graphs, and pixel-based imaging approaches. These methods are analyzed based on their ability to balance invertibility with symmetry preservation. With respect to generative model architectures, we trace the technological evolution from latent space-based models, such as Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN), to state-of-the-art frameworks including diffusion models, flow-matching, and Transformer-based autoregressive models (e.g., MatterGen, MatterGPT). The review highlights how these architectures integrate physical constraints to navigate the vast chemical space of crystalline materials efficiently