[AI Journey track] GenAI: from digital to physical modalities

The advent of GenAI has ushered in an era characterized by multitasking, multiagency, and multimodality. While multimodality has become increasingly crucial in AI development, with models based on both monolithic and distributed architectures capable of synthesizing and describing text, images, and sound, the potential of GenAI extends far beyond these common digital modalities. This presentation explores the expanding frontier of GenAI applications in the physical sciences, particularly in solid-state physics and molecular physics. We demonstrate how, with appropriately selected representations of physical objects, the same neural network architectures underlying GenAI (such as transformers and diffusion models) can be adapted to predict and synthesize novel materials and structures. Our talk will showcase recent breakthroughs in this field, presenting examples of AI-predicted physical objects that have been successfully brought to experimental synthesis. We will address the core challenges in applying GenAI to the physical world, including:- • The strict requirement to respect physical laws and principles, necessitating validation through physics-based numerical solvers (e.g., density functional theory in solid-state physics) • The need for sufficient data volume to train foundation models in specialized scientific domains. Furthermore, we will discuss the potential implications of these advancements for accelerating scientific discovery and materials design. This presentation aims to stimulate interdisciplinary dialogue and highlight the transformative potential of GenAI in bridging the digital and material realms of scientific research.