In January, Qiyuan Chen presented MRSEC-supported research at the 40th AAAI Conference on Artificial Intelligence (AAAI-26) in Singapore. With a highly selective acceptance rate of 17.6% this year, AAAI remains one of the premier international venues for peer-reviewed research in artificial intelligence (AI). Chen’s presentation introduced a new physics-aware generative AI framework designed to decode the complex structure of disordered materials.
Generative AI Meets Disordered Materials
Generative AI has rapidly gained attention in materials science. However, disordered materials pose unique challenges. Unlike crystalline materials, disordered materials with the same chemical composition can exhibit an astronomically large number of possible atomic structures. While these structures share common structural motifs, they also contain subtle variations that can lead to different behaviors and properties.
Standard generative models, which typically function as “black boxes,” often produce mathematically plausible atomic structures that are nevertheless physically unrealistic. For example, generated structures may be energetically unstable or fail to reproduce the key structural motifs and signatures observed in real disordered materials.
A Physics-Regularized Hierarchical Generative Model
Enabling generative AI for materials like metallic glasses first requires an effective numerical representation of disordered atomic structures. Such a representation must detect subtle variations within complex atomic networks, even if the structure is shifted, rotated, or described in different orderings of atoms. At the same time, it must be computationally efficient while capturing structural features most relevant to a material’s energy and behavior.
To meet these challenges, Chen and collaborators of Bu Wang’s lab developed a physical-regularized hierarchical generative model designed specifically for disordered metallic systems. The model builds on advanced AI architectures but is carefully designed so that its predictions depend only on meaningful structural differences, not on how the structure is positioned, oriented, or labeled. Beyond this design, the team incorporated physics-based guidance during training. They introduced constraints based on structural patterns known to be characteristic of disordered materials, along with structural energy information. Together, these serve as “physics guardrails,” discouraging the model from generating unstable or unrealistic atomic configurations. As a result, the structures produced by the model are not only statistically representative of real data but also consistent with fundamental physical principles.
Mapping Atomic Dynamics in Disordered Systems
A key outcome of this approach is the creation of a physics-aware “latent space”—a simplified map that captures the underlying energy landscape of metallic glasses. Rather than relying solely on large-scale molecular dynamics simulations or computationally demanding supercomputer calculations, researchers can explore and sample this learned map to study how disordered atomic structures evolve, rearrange, and settle into stable forms.
This capability is part of the MRSEC’s broader efforts to understand the dynamics in disordered material systems. By combining generative AI with embedded physical principles, the framework represents a shift from purely data-driven modeling toward materials AI that is more interpretable, physically grounded, and scientifically insightful. The full research article can be accessed through https://arxiv.org/pdf/2505.09977