CAREER: Generative 4D Simulation with Physical Constraints

This project aims to improve artificial intelligence (AI) systems that predict how three-dimensional environments change over time. Current AI models that generate videos or simulate future events often require massive computing power and routinely violate basic laws of physics, such as allowing solid objects to pass through one another. These flaws make them unsafe for critical real-world applications like medical surgery, robotics, or scientific simulation of future events. This award supports the development of a new type of artificial intelligence that creates fast and physically realistic simulations of moving environments. Building AI systems that respect physical laws will strengthen the United States' AI-powered scientific leadership and advance the national health by enabling safer medical treatments. Specifically, the project applies these new AI-based simulations to cancer radiation therapy, aiming to predict real-time tumor movements so that treatments can precisely target the cancer while minimizing harm to healthy tissue. The project also supports educational activities, including creating a new university course on computer graphics at the University of Michigan, mentoring undergraduate researchers, and hosting summer camps for high school students to learn about artificial intelligence. This award advances the fields of generative modeling and physical simulation by developing conditional sparse neural fields for four-dimensional scene generation. To overcome the computational bottlenecks of dense spatial sampling in current diffusion models, the investigator will design autoregressive networks that query only a sparse subset of points while adaptively maintaining full fidelity on regions of high physical importance. To ensure physical plausibility, the research integrates physical constraints typically described as partial differential equations (PDEs), such as momentum conservation, topology preservation, and incompressibility, directly into the generative models. Rather than relying on computationally expensive test-time optimization, these constraints are enforced through efficient residual conditioning during both training and inference. Finally, the project translates these advancements to medical imaging by developing a real-time predictive framework for organ motion. Using sparse clinical observations, such as magnetic resonance imaging slices, the system will forecast tumor dynamics to enable predictive, rather than reactive, adaptive radiation therapy for precise targeting. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2544116 | Program: 01002930DB NSF RESEARCH & RELATED ACTIVIT,01003031DB NSF RESEARCH & RELATED ACTIVIT,01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Jeong Joon Park | Institution: Regents of the University of Michigan - Ann Arbor, ANN ARBOR, MI | Award Amount: $367,296 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2544116 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2544116.html