Physics-informed machine learning and neural rendering for inverse problems, differentiable physics, dynamic inference, and scientific simulation.
We design computational models that bridge inverse problems, machine learning, and differentiable physics to reveal hidden structure and dynamics from complex sensor observations
Differentiable physics enables computing gradients through the entire sensing pipeline — from the imaging system to the physical processes that generate the observations, including light–matter interactions, scene dynamics, and even the curvature of space-time.
Neural representations offer a powerful way to model high-dimensional continuous fields. We develop neural representations, with physics-guided inductive biases, for reconstructing the stucture and dynamics of physical fields and instrument behavior from sparse or indirect observations.
Learning dynamical systems and temporal evolution directly from sparse, noisy, or indirect data using models that capture multi-scale physical processes and underlying latent structure.
Learning fast, differentiable surrogates for expensive physical simulations, enabling rapid scientific modeling, inference, and exploration across complex physical systems.
