Research Theme

Generative Physical System Emulation

We design generative emulators of physical systems that learn minimal, closed representations of high-dimensional dynamics— enabling fast simulation, system identification, and physically interpretable inference.

Reduced-Order Modeling Data-Driven Closure Physics Constraints

Why generative physical system emulation?

Many physical systems—climate, turbulence, sediment transport, energy systems—are observed in extremely high dimensions, yet evolve on much lower-dimensional structures.

Traditional surrogates focus on reproducing outputs.
Our goal is to emulate the system itself.

The central challenge is to learn the minimal state representation and the closed dynamics that evolve it—while preserving physical structure.

In this view, uncertainty is not an afterthought: it reflects unresolved scales, regime dependence, and structural ambiguity that must be embedded in the emulator itself.

A framework for minimal and closed representations

🧭 Minimal dimension discovery

Learn reduced coordinates that capture the system’s attractor or manifold while retaining the mechanisms that govern evolution.

🧩 Closure learning

Replace unresolved processes with data-driven closures that remain stable, interpretable, and physically consistent.

⚡ Fast operator emulation

Neural operators and generative flows enable millisecond-scale inference for ensembles, sensitivity analysis, and scenario exploration.

What this framework enables

Reduced-order system emulators that remain closed under evolution
Validated physical closures with uncertainty bounds
Regime-aware simulation, including extremes and rare transitions
Real-time ensemble generation for stress testing and decision support
Interpretable system identification, not black-box prediction

Science → operations

Generative emulators provide a pathway from fundamental modeling to operational use.

Our focus is on transition-ready emulators: minimal, validated, and physically interpretable surrogates that reduce computational cost by orders of magnitude while retaining scientific meaning.

Outcome

A unified platform for learning, emulating, and understanding complex physical systems— revealing minimal structure, enabling fast simulation, and supporting decision-grade analysis.