My research develops the principles of a general “science of foundation models,” with the overarching goal of developing powerful scientific research agents. I focus on three interconnected areas:
- understanding the end-to-end foundation model design pipeline from scientific principles,
- developing data efficient foundation model training and adaptation techniques, and
- adapting foundation model pipelines to non-standard scientific domains.
Through scaling laws research with Meta, I have shown that knowledge acquisition and reasoning tasks require fundamentally different compute allocation strategies—challenging the assumption of uniform scaling. I complement this with automated ML techniques that adapt general models to domains like protein folding and climate modeling, where labeled data is scarce and expert knowledge is critical. My goal is enabling scientists to leverage state-of-the-art AI without requiring extensive ML expertise, accelerating discovery across fields from biology to physics.