Scaling Scientific Discovery: Accelerating the Discovery Loop

In this talk on scaling scientific discovery, Felix Schuermann of the Google Cloud HPC team, formerly co-director of EPFL's Blue Brain Project, traces how the discovery loop is being accelerated by AI. He recalls two decades of simulation-driven neuroscience enabled by Moore's law, then contrasts it with protein structure prediction, where brute-force compute failed and DeepMind's AlphaFold succeeded by learning from known structures, with AlphaFold 3 extending to protein-DNA, RNA, and ligand interactions. He positions bespoke science AI models such as WeatherNext and AlphaGenome as a new path to prediction alongside experiment, theory, and simulation. The talk then turns to agentic systems: AlphaEvolve, an LLM-driven evolutionary optimizer that rewrites entire programs against a formal evaluation function, improved Google's fleet scheduler by nearly one percent and found more power-efficient matrix-multiplication circuits for chip design. Multi-agent frameworks such as the CoScientist use tournament-style ranking and debating agents to generate and refine research hypotheses grounded in a lab's own data. He outlines the infrastructure implications: strong reasoning LLMs with large context windows, data grounding with privacy guarantees, elastic bursts of inference calls, and many CPU sandboxes for agent tools. In the Q&A he addresses accelerating physical experimentation through automated labs and keeping humans in control of costs and final decisions.