Refreshments are available starting at 10:30 a.m. The seminar will begin at 10:45 a.m.
Abstract
Scientific discovery has long been constrained by the exponential expansion of hypothesis spaces and the prohibitive cost of validating hypotheses through high-fidelity simulations or experiments. In this talk, Yuanqi Du will present a unified framework for AI-driven scientific discovery organized around three pillars. First, he will discuss how generative models trained on scientific data can serve as flexible, data-driven priors over structured hypothesis spaces, with inference-time steering enabling adaptation to diverse downstream tasks without retraining. Second, Du will unpack the principled foundation underlying this flexibility: a fundamental connection between non-equilibrium thermodynamics and probabilistic inference that not only provides tools for steering generative models, but also opens up new avenues for scalable molecular simulation. Third, he will show how the growing scientific knowledge embedded in large language models opens up new possibilities for automating the discovery workflow and closing the search-and-validate loop across scientific domains. Together, these three pillars chart a path toward meaningfully accelerating the rate of scientific discovery.
Speaker Biography
Yuanqi Du is a PhD candidate in Computer Science at Cornell University. His research focuses on developing principled and efficient probabilistic and geometric modeling methods that are inspired by, and accelerate, discovery in the natural sciences, spanning chemistry, physics, and biology. His work has appeared at leading machine learning venues (e.g., the Conference on Neural Information Processing Systems, the International Conference on Machine Learning, the International Conference on Learning Representations) and in scientific journals including Nature, Nature Machine Intelligence, Nature Computational Science, and the Journal of the American Chemical Society, including three cover articles. As a passionate community builder, Du has organized over 20 events including conferences, workshops, and seminar series on topics such as AI for science, probabilistic machine learning, and learning on graphs.