Research projects
One logic across the projects
Each project starts with a physical quantity to predict or reconstruct and the inputs, geometry, constraints, and disturbances that affect how it changes. A learned model estimates that quantity, which is then used to compute a control or optimization decision. Evaluation measures how prediction errors affect that decision. The projects are grouped by the three directions, and each description identifies the system used for evaluation.
- Physical state and inputs
- Dynamics and uncertainty
- Learned solution or forecast
- Control or optimization
- Decision-level evaluation
A. Physical dynamics in science and engineering
iCVD coupled with liquid crystals Science
With Professor Fengqi You at Cornell, my current work focuses on initiated chemical vapor deposition (iCVD) coupled with liquid crystals (LC). I formulate equations describing this process and train neural models to approximate their solutions for control and optimization. Chemical formation coupled with LC is a possible extension.
Power, water, and industrial processes Engineering
During my Ph.D. with Professor Jingcheng Wang at Shanghai Jiao Tong University, I worked on power generation, urban water supply, and hot-strip rolling. Models of these processes relate the operating inputs to pressure, temperature, and energy consumption. Demand and electricity price are external inputs to the operating decisions. The projects under B and C were tested on these systems.
B. Neural solvers for physical dynamics
Repeated control and optimization require the governing model to be solved for many inputs and conditions. I develop neural solvers, including neural operators, that approximate the solution map at a lower cost. The current questions are the accuracy of the predicted trajectory, the accuracy of its sensitivities to process inputs, and the gain from equation constraints and hybrid numerical structure.
Temporal response and uncertain pressure states Engineering
Daily demand and network pressure expose different errors in a learned model. KbBcM treats the delayed response near a demand turning point as a timing error and corrects it with lag-penalty attention and a residual step. CritiCoder separates the pressure response to measured inputs from the variation caused by unmeasured disturbances. Both studies report response timing and disturbance sensitivity next to the average error.
Multi-step energy dynamics in industrial rolling Engineering
Hot-strip roughing is an ordered sequence of rolling passes, so the error of one prediction is carried into the next. A DCLS-Net study predicts the energy of each pass, and series concatenation links the passes so that the error stops accumulating along the sequence. The prediction object is the process trajectory, not an isolated sample.
C. Model2Action
The consequences of a solver error depend on how the solution is used in control or optimization. I study when a change in the predicted dynamics changes the selected decision, and when that change increases cost or violates a constraint. Regimes near a bifurcation or a stability boundary, and points where the set of active constraints changes, are the main cases. The measured decision error then serves as the training signal for the solver.
Coupled water-energy operation under forecast uncertainty Engineering
Urban water treatment, transport, and distribution consume electricity, so an uncertain electricity price can change the preferred schedule. SPTO provides a forecast distribution to a stochastic water-supply model and trains the forecast through a duality-derived upper bound on the optimality gap of the schedule. Operating cost and regret measure how forecast errors affect the decision. This is my first study to quantify decision error.
Feedback laws for constrained nonlinear dynamics Engineering
Two last-author studies concern learned feedback laws for nonlinear dynamics. For discrete-time constrained systems, nonlinear model predictive control initializes a controller network, and a control barrier function with a nonquadratic loss enforces the state and input constraints during learning. For adaptive optimal control, eligibility traces at several time scales give the critic credit signals accumulated over different horizons.
Constrained controller · Multiple time-scale eligibility traces
Geometry-conditioned motion in physical environments Robotics
DRIFT stores local LiDAR geometry as graph memory and retrieves waypoint-specific context during recurrent diffusion. Endpoint error, clearance, smoothness, and runtime measure how the generated trajectory relates to the target and to the local geometry. The question is the same as in direction C, whether a learned action respects the constraints of the environment it acts in.
The publications are listed by year with the same tags.