FORWARD →
What would this look like?
Simulate, predict, and generate. A learned model connects underlying structure and dynamics to the observations we can measure.
Research
We build world models—generative, predictive models of how things look, move, and evolve—and use them to solve inference problems in the physical and living world.
01 / THE MODELING CORE
Generative & world models
We study diffusion, flow, video, and 3D models that can predict and generate complex systems. We design conditioning interfaces that go beyond text: measurements, geometry, physics, and clinical variables.
Diffusion & flow modelsVideo & 3D generationControllable generation
02 / THE PHYSICAL WORLD
Models for science
We develop learned operators and simulators for physical systems, and solve inverse problems using generative priors. Our interests include computational imaging, PDE surrogates, inverse design, and active sensing.
Operator learningInverse problemsComputational imaging
03 / THE LIVING WORLD
Models for health
We build multimodal models across medical images, clinical text, and physiological signals. We study temporal imaging, diagnosis and prognosis, procedural video, and decision support with calibrated uncertainty.
Multimodal learningMedical imagingReliable clinical AI
OUR ORGANIZING IDEA
Forward and inverse.
Two sides of the same model.
← INVERSE
What produced this?
Infer from partial, noisy measurements. The same model becomes a prior for reconstruction, with uncertainty as part of the answer.
Our methodological foundation spans generative modeling, operator learning, inference, and representation learning. Efficiency and reliable evaluation run through all of these areas.
QUESTIONS WE WORK ON
Research directions
One model, many sensors
Resolution- and sampling-agnostic reconstruction across imaging modalities and acquisition settings.
Temporal & 3D models for medicine
Dynamic ultrasound, 4D imaging, and physiological dynamics as predictive models, beyond static snapshots.
Conditioning beyond text
Specifying generation through measurements, geometry, physical constraints, and clinical variables.
Deployment-aligned optimization
Learning and evaluation centered on ranking, calibration, and cost under class imbalance.
Learned simulators for inverse design
Fast, differentiable forward models that make difficult inverse problems tractable.
Reliable AI assistants
Tool-using, multimodal systems for science and clinical workflows, with uncertainty-aware escalation.