Dustin Sell is a Technical Specialist and Customer Engineer with Google Cloud Public Sector, specializing in artificial intelligence architectures, enterprise deployment, and AI-driven scientific enablement across the U.S. Department of Energy (DOE) and National Laboratories. In his role, Dustin collaborates with lab researchers, leadership, and IT teams to implement secure, scalable cloud environments and frontier AI tools—including Gemini Enterprise and Google DeepMind’s AI platform. He focuses on bridging cutting-edge generative AI research with autonomous scientific workflows to empower national lab scientists in accelerating discoveries across high-impact research domains.
Presentation Title:
Gemini for Science: Transforming Discovery Through Multi-Agent Causal Reasoning and Autonomous Hypothesis Generation
Presentation Abstract:
Accelerating the pace of scientific discovery requires moving beyond traditional machine learning paradigms that focus solely on narrow parameter optimization, curve fitting, or passive literature retrieval. To address complex grand challenges across materials science, critical minerals exploration, and climate dynamics, researchers require intelligent systems capable of active reasoning over high-dimensional scientific domain knowledge, discovering non-obvious causal links, and proposing novel, testable hypotheses.
This presentation and tutorial introduces Gemini for Science and its frontier AI Co-Scientist architecture. Built upon Google’s multimodal Gemini models, Gemini for Science mirrors the scientific method through a specialized multi-agent framework comprising dynamic Generation, Reflection, Ranking, Evolution, Proximity, and Meta-Review agents. Rather than acting merely as chatbots or simple summarizers, these autonomous agents collaborate to synthesize complex literature, reason over physical and chemical mechanisms, formulate original scientific hypotheses, and design actionable experimental protocols.
During this 40-minute tutorial and demonstration session, we will walk through the core architecture of Gemini for Science and showcase real-world agentic workflows applied to scientific challenges:
- Multi-Agent Orchestration & Evolution: How automated feedback loops iteratively critique, evolve, and rank scientific hypotheses to elevate novelty and rigor.
- Reasoning Over Causal Chains: How agents evaluate mechanistic pathways and causal dependencies rather than simple statistical correlations.
- Human-AI Scientific Collaboration: Methods for researchers to steer agent behavior through natural language goals, seed ideas, and interactive constraint settings.
- National Lab Integration: Practical considerations for deploying agentic architectures alongside high-performance computing (HPC) workflows, digital twins, and secure cloud environments.
Attendees will leave with a clear understanding of how agentic AI systems can serve as virtual co-scientists to expand human cognitive throughput and accelerate scientific breakthroughs across DOE National Laboratories and the broader scientific community.