The Physics Colloquium has been arranged regularly since the fall semester of 2021. The detailed schedule and talk information of this semester are as follows.
For inquiries or suggestions of future speakers, please contact the colloquium working group (Prof. Jane Lixin Dai, Prof. Tran Trung Luu, Prof. Yanjun Tu, Prof. Chenjie Wang, and Prof. Shizhong Zhang).
September 23, 2026 (Wed) 3:30 p.m.
September 30, 2026 (Wed) 4:00 p.m.
October 7, 2026 (Wed) 4:00 p.m.
October 21, 2026 (Wed) 10:30 a.m.
November 4, 2026 (Wed) TBC
November 18, 2026 (Wed) 10:30 a.m.
December 2, 2026 (Wed) 4:00 p.m.
Can AI Turn a Physics Paper into a Validated Scientific Program?
Lessons from Quantum Many-Body Programming
Speaker: Prof. Yi ZHOU
Affiliation: Institute of Physics, Chinese Academy of Sciences
Date: September 23, 2026 (Wed)
Time: 3:30 p.m.
Venue: MB237, 2/F, Main Building, Main Campus, HKU
Poster: TBC
Abstract:
Can an AI read a physics paper and turn it into trustworthy research code? In quantum many-body physics, papers often leave implicit the index ordering, gauge choices, fermionic signs, truncation rules, contraction strategy, and scaling decisions that determine whether an implementation is correct.
In this colloquium, I will present a human-in-the-loop “Paper-to-Program Many-Body” workflow that externalizes this hidden knowledge before coding. A virtual research group of AI agents extracts the theory, reviews an implementation-ready specification, and generates code tested by physics oracles and production-scale checks.
Two case studies show promise and limits. For DMRG, specification-guided implementations succeeded in 16 of 16 tested pairings, versus 6 of 13 direct attempts. For the sign-sensitive Pfaffian conversion of paired-fermion states to matrix product states, the workflow produced 11 validated passes among 26 audited runs, versus none under direct prompting after provenance and scale auditing. The remaining failures expose a second bottleneck: even with a good specification, current agents differ in their ability to reason, debug, and construct scalable scientific software.
The broader lesson is that reliable AI-assisted science requires more than better code generation. It requires explicit knowledge externalization, expert review, and validation protocols that make both successes and failures auditable.
Date: September 30, 2026 (Wed)
Time: 4:00 p.m.
Venue: MWT5, 1/F, Meng Wah Complex, Main Campus, HKU
Date: October 7, 2026 (Wed)
Time: 4:00 p.m.
Venue: MWT5, 1/F, Meng Wah Complex, Main Campus, HKU
Date: October 21, 2026 (Wed)
Time: 10:30 a.m.
Venue: MB201, 1/F, Main Building, Main Campus, HKU
Date: November 4, 2026 (Wed)
Time: TBC
Venue: TBC
Date: November 18, 2026 (Wed)
Time: 10:30 a.m.
Venue: KKLG109, LG1/F, K.K. Leung Building, Main Campus, HKU
Date: December 2, 2026 (Wed)
Time: 4:00 p.m.
Venue: CYCP1, LG1/F, Chong Yuet Ming Chemistry Building, Main Campus, HKU