Towards Agentic AI for Science Hypothesis Generation, Comprehension, Quantification, and Validation

Abstract

AI is revolutionizing scientific discovery by connecting seemingly unrelated fields – from mechanics to biology, and science to art. However, how can we build AI models that don’t merely retrieve information but make new discoveries, going beyond interpolation to extrapolate to reason over never-beforeseen scenarios and concepts? In this talk we describe how a new generation of physics-aware AI is breaking traditional boundaries through: • Innovative graph-based generative AI combining physics and data-driven modeling • Biologically-inspired neural structures that adapt dynamically • Multi-agent systems that mirror natural systems Through practical case studies, I will present how this technology transforms materials science across scales – from silk and collagen to biomineralized materials – with direct applications in medicine, food systems, and agriculture. The versatility in agent development allows for expertise in diverse domains, including knowledge retrieval, protein structure analysis, physics-based simulations, and results analysis, is presented. The dynamic collaboration between agents, empowered by LLMs that can reason over sequences, data, images, and text, provides a versatile approach to tackling protein design and analysis problems, as demonstrated through diverse examples in this study.