AI for Physics
Structure-aware learning, interatomic potentials, and reliable physical simulation.
Taoyong Cui · 崔涛镛
Building intelligence for scientific discovery.
Toward reliable physical intelligence, world models, and autonomous scientific discovery.
I am a Ph.D. student in the Department of Computer Science and Engineering at The Chinese University of Hong Kong (CUHK), working with the AI4LS Laboratory and Multimedia Laboratory (MMLab). I am also a visiting Ph.D. researcher at Stanford University.
From atoms to agents Reliable learning systems grounded in scientific structure.
About
Research at the intersection of intelligence and the physical world.
I develop scientifically grounded AI that can reason about complex physical systems, adapt beyond its training distribution, and collaborate in open-ended discovery.
At CUHK, I conduct research with the AI4LS Laboratory and MMLab, co-supervised by Prof. Pheng Ann Heng and Prof. Wanli Ouyang. At Stanford University, I collaborate with Prof. Le Cong. Before CUHK, I earned an Academic Master's degree in Biomedical Engineering from Tsinghua University and completed research internships at Microsoft Research Asia and Shanghai AI Laboratory.
I have served as a reviewer for Nature Communications, AISTATS, ICML, ICLR, NeurIPS, and AAAI. I welcome conversations around scientific machine learning and collaborative research.
World Models
Representations that learn dynamics, uncertainty, and generalizable causal structure.
Agentic Systems
Adaptive agents that plan, use tools, and participate in iterative scientific workflows.
Selected research
Reliable learning for molecular and physical systems.
Selected work on uncertainty, adaptation, and geometric representation learning for machine-learned interatomic potentials.
Nature Communications · 2025
Evidential deep learning for interatomic potentials
A physics-inspired evidential framework that quantifies uncertainty without significant computational overhead or reduced prediction accuracy.
Read paperNature Communications · 2025
Online test-time adaptation for better generalization of interatomic potentials to out-of-distribution data
TAIP uses dual-level self-supervision across global structures and local atomic environments to adapt interatomic potentials at test time.
Read paperNature Machine Intelligence · 2024
Geometry-enhanced pretraining on interatomic potentials
A two-stage framework combines molecular dynamics data generation with masking, denoising, and contrastive learning to capture geometric structure.
WAIC 2024 Outstanding Thesis Award · GBA AI for Science Ph.D. Forum Best Paper
Read paper News
Recent highlights.
Publication updates and milestones from my ongoing research.
A paper was accepted by Nature Communications.
A paper was accepted by Scientific Data.
A paper was accepted by Current Opinion in Structural Biology.
A paper was accepted by Advanced Science.
A paper was accepted by Nature Communications.
A paper was accepted by VLDB.
A paper was accepted by Nature Machine Intelligence.
Talks
Sharing ideas across AI and science.
Global Artificial Intelligence Technology Conference
ConferenceGuangdong–Hong Kong–Macao Greater Bay Area AI for Science Ph.D. Forum
Research forumWorld Artificial Intelligence Conference
ConferenceChina Academic Forum on Interdisciplinary Innovation for Graduate Students in Materials Science
Academic forumGuangdong–Hong Kong–Macao Greater Bay Area AI for Science Ph.D. Forum
Research forumWorld Artificial Intelligence Conference
ConferenceContact