Talk as part of a colloquium in Data-Driven Materials discovery

Summary: Modern AI agents for scientific discovery face two challenges: poor out-of-distribution (OOD) reliability and prohibitive computational and sample costs. We address these through a holistic approach spanning optimization algorithms, neural architectures, and specialized hardware. Algorithmically, we bridge uncertainty quantification and policy optimization to achieve state-of-the-art sample efficiency in protein search and therapeutic molecular design. Architecturally, we incorporate expressive state-space models to reduce computational complexity while improving OOD generalization in diverse state-tracking tasks. At the hardware level, we map these primitives onto emerging in-memory computing hardware, exploiting physical device properties for extreme energy efficiency. Together, these advances pave the way for efficient and reliable AI agents for autonomous scientific discovery.