Research
Physical intelligence and system dynamics
Physical intelligence is the ability of a learning system to model a physical system and to act on it. A material during processing, a water network under pumping, and a power plant responding to load changes evolve according to ordinary or partial differential equations with initial conditions, boundary conditions, and external forcing. To model such a system is to solve those equations under the given conditions. To act on it is to compute a decision from the solution. I study both steps and how solution errors affect decisions. I organize this work into three research directions.
A. Physical dynamics in science and engineering
A governing model specifies the physical state, its evolution under inputs and forcing, and the conditions that constrain it. Ordinary differential equations describe lumped dynamics. Partial differential equations describe fields distributed in space. With initial and boundary conditions, these equations define the solution problems of direction B.
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.
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. Their states are pressure, temperature, and energy consumption. Demand and electricity price are external inputs to their operating decisions.
B. Neural solvers for physical dynamics
Repeated control and optimization require the governing model to be solved for many inputs and conditions. A neural solver, such as a neural operator, approximates the solution map across a family of conditions and returns a solution in one forward pass. I train these models so that their rollouts remain accurate over the prediction horizon required by the controller and their sensitivities to process settings are accurate enough for an optimizer. Physics-constrained training and hybrid schemes that keep part of the numerical solver reduce the data they need. Computational cost and data requirements are evaluated together with accuracy.
My earlier engineering studies dealt with the same errors in learned temporal models. Lag-free demand forecasting reduces the delayed response of a demand model near turning points. Uncertain pressure regression separates the pressure response to measured inputs from the variation caused by unmeasured disturbances. A series-concatenation model of hot-strip rolling links successive passes so that the error stops accumulating along the sequence.
C. Model2Action
Once a learned solution is used to compute a decision, its error affects the result through the dynamics, the objective, and the constraints. I separate the change in the selected decision from its consequences. Two substantially different decisions can have nearly the same cost, so a large change in the decision can still correspond to low regret. Closed-loop cost and feasibility need their own evaluation.
The dependence on the operating regime is the main question. Near a bifurcation or a stability boundary, a small model error can change the predicted trajectory or its stability. When the set of active constraints changes, the response of the optimizer changes with it. I locate the regimes where these effects are large, and I use the measured decision error to train the solver.
Stochastic Predict-Then-Optimize is an earlier example of training through the decision. It trains a probabilistic electricity-cost forecast through the water-supply schedule the forecast induces, and its SPTO+ variant uses a differentiable upper bound on the optimality gap as the loss. Two control studies concern the decisions themselves. A reinforcement learning controller for constrained nonlinear systems enforces state constraints during learning with a control barrier function. Adaptive optimal control with multiple time-scale eligibility traces updates the critic with information accumulated over several horizons. DRIFT conditions recurrent diffusion on local LiDAR geometry, so a generated trajectory stays consistent with the space it moves through.
Research map
| Direction | Question | Evaluated by |
|---|---|---|
| A. Physical dynamics in science and engineering | Which state evolves, under which inputs and conditions? | Consistency with the governing relations and with observations |
| B. Neural solvers for physical dynamics | How fast and how accurately can the governing model be solved? | Solution error, rollout error, sensitivity error, and computational cost |
| C. Model2Action | When does model error change a decision, and by how much? | Decision change, regret, closed-loop cost, and feasibility |
Publications are listed by year with a tag for each direction they belong to. Research projects are grouped by direction.