CHENGJIE FANG / RESEARCH ARCHIVEUCAS 2026.11

03 / Research programs

Research directions

I study learning systems that use physical structure, task representations, and feedback to model the world and act within it.

01

Physics-aware learning / Spatiotemporal modeling

Scientific machine learning

My work on ocean carbon dynamics studies how learning systems can represent the different processes that shape a complex physical system. Carbon-ECS combines transport modeling with local detail refinement. REM-MoE uses hierarchical experts to separate environmental regimes from physical mechanisms.

My current work also investigates how temporal reaction search can be separated from spatial physical response for ocean carbon-flux forecasting.

02

Robot navigation / Reinforcement learning

Embodied intelligence

I study how agents represent task progress, verify intermediate states, and recover from failure in long-horizon navigation. StageNav explores geometric verification and state-aligned reinforcement learning, while ForesightNav combines semantic understanding with geometric exploration.

Earlier work / Undergraduate robotics

Humanoid perception and robotic manipulation

At Wuhan Institute of Technology, I worked with Tongzhou Zhao and Xun Li. I joined the university's robotics laboratory in my second year, working on image processing and control for humanoid robots.

In my third year, I led a robotic-arm project for board-game playing using deep learning and reinforcement learning. The project was entered in national student competitions. These experiences shaped my interest in the connection between perception, decision-making, and robot behavior.

Undergraduate project background
03

Autonomous planning / Scientific experimentation

Autonomous research agents

I build systems that coordinate planning, context-state management, experimentation, and model refinement. My internships at Baidu and ByteDance connect these ideas to scientific workflows and forecasting.

My current manuscripts investigate scientific-state representations for autonomous PDE modeling and structured ocean prediction. Both remain under review.

Research agents

UNDER REVIEWFirst author

What Models Miss: Scientific State Representations for Autonomous PDE Modeling

Scientific state representations and residual-driven exploration for autonomous PDE model search and refinement.

Scientific AI

UNDER REVIEWFirst author

Decoupling Temporal Reaction Search from Potential-Wave Spatial Response

Separating temporal reaction search from differentiable spatial physics for ocean carbon-flux forecasting.

Industry research experience

04

Supporting work / Computer vision

Visual perception

My collaborative work also covers lightweight defect detection, remote-sensing forgery perception, and zero-shot template matching. These projects explore visual representations across different scales and imaging conditions.

Research figure