Title: Generative Models of Motif-Hierarchies in Complex, Disordered Materials

Understanding complex materials is challenging because of the intricate interplay of numerous components and emergent behaviors. These materials exhibit dynamic interactions that span many length and time scales, giving rise to non-equilibrium phases, functional disorder, and hierarchical self-organization.

Modeling complex materials is arduous due to the vast number of degrees of freedom that interact non-linearly. Yet at the heart of these materials are structural motifs -- recurring atomic arrangements that serve as fundamental building blocks. These motifs not only dictate the local structure but also govern the emergence of larger mesoscale and macroscale features. Their interactions, especially in disordered systems, influence transitions between order and disorder, packing density, and material properties. Despite their importance, the relationships between structural motifs and their role in shaping complex materials are not fully understood.

Structural motifs assemble into larger units through complex rules, much like words form phrases and narratives in language. Using concepts from language modeling, generative models can predict intermediate motifs and contextualize them into the larger structures that define boundaries and domains.

In this talk, I will highlight how machine learning has been applied to uncover the hidden language of structural motifs in complex materials such as piezoelectrics. By combining Markov models for short-range interactions with coarse-grained nucleation models for longer-range phenomena, we captured the disorder and complexity of these systems with remarkable efficiency. These three components map to the concepts of vocabulary, grammar, and context in language machine models.

I will also briefly explain how the learning of such motif hierarchies can help us tackle complex interactions in active matter, supercooled water, and even bio-imaging and detection.