Bio

Zhe Fu is a final-year Ph.D. candidate in Transportation Engineering and M.S. candidate in EECS at the University of California, Berkeley, advised by Prof. Alexandre Bayen at BAIR and Berkeley ITS. Her research focuses on learning, control, and modeling for distributed parameter systems, with an emphasis on mixed autonomy. She develops physics-informed neural models of hyperbolic PDEs, designs both model-based and data-driven control algorithms, and validates them through large-scale field experiments. Zhe has been recognized as a 2025 Eno Fellow and was the Runner-up in the 2025 Berkeley Grad Slam. Her research has received honors across communities, including First Place in the INFORMS Poster Competition (2023) and Rising Stars awards in Mechanical Engineering (2024), EECS (2025), and NSF CPS (2025). Her leadership, mentorship, and teaching efforts have been recognized by UC Berkeley and organizations such as ITS/CTF, EDGE in Tech, H2H8 and AAa/e. Learn more at https://fu-zhe.com/

Physics-Informed Learning and Control for Mixed Autonomy Traffic Systems

Mixed autonomy traffic systems, where automated and human-driven vehicles coexist, represent one of the most complex and high-impact challenges in intelligent transportation. These systems are dynamic, multi-agent, and safety-critical, demanding new methods that can model and influence their behavior under uncertainty and real-world constraints.

My research develops physics-informed learning and control algorithms to address three core challenges: 1) capturing the dynamics of large-scale systems where traditional models often fall short; 2) designing control algorithms that steer system behavior with minimal intervention; and 3) ensuring that these approaches remain safe and robust outside simulation. I focus on mobility systems because they sit at the intersection of algorithmic complexity and societal impact.

My work begins with modeling. I developed neural finite-volume methods for hyperbolic partial differential equations (PDEs) that preserve conservation and entropy while improving predictive accuracy and efficiency over classical numerical schemes. These models serve as the backbone for real-time control. On the control side, I designed kernel-based and imitation learning algorithms to guide mixed-autonomy traffic using a small fraction of automated vehicles, validated both in simulation and in the field. I played a key role in the first mixed autonomy large operational field test, where 100 AVs were deployed on the I-24 highway during the morning rush, demonstrating fuel savings and improved flow. This experience highlighted the need for formal safety guarantees. I developed neural control barrier functions that bridge model-based safety certification with learning-based systems, particularly in safety-critical platforms like drones and urban air mobility.

Looking forward, I aim to extend these methods to higher-dimensional systems with stronger coupling and real-world constraints, toward building intelligent mobility systems that are sustainable, equitable, and deployment-ready.