Towards AI Systems for Scientific Discovery

This talk will begin by introducing ideas towards building a comprehensive framework for AI-augmented scientific discovery combining large reasoning models with domain-specific foundation models that share embedding spaces and interfacing them with dynamic knowledge graphs and AI agents that perform a number of tasks including formal verification and managing interactions with the real world via simulations and experiments. Following this, specific instantiations of different elements of this AI system will be demonstrated, including large scale foundation models and agents for molecular discovery and spatio-temporal physics. The potential of this technology and wide-ranging gaps will be highlighted.