CIRC: New: COSTA: An Open Cloud Simulation Platform for Physical Intelligence

Currently it is costly and difficult to simulate complicated physical scenarios such as tsunamis, earthquakes, or explosions. The mathematical and physical models that describe these effects are highly complex, and solving these models to produce realistic simulations can often require tremendous computing power. Yet physical simulations and digital twins are becoming increasingly crucial tools for researchers in the NSF Directorate for Computer and Information Science and Engineering (CISE), especially as artificial intelligence (AI) models are starting to interact with the real world. Whether considering models that power self-driving cars, household robots, or industrial design tools, a software platform for "physical intelligence" could allow researchers to create a new generation of innovations guided by mathematics and physics. In light of this, this collaborative project brings together investigators from Vanderbilt University, Georgia Institute of Technology, University of California Davis, and Stanford University to create a new, sustainable, community-driven software platform for physical intelligence. The project consists of several main thrusts to build the proposed platform, COSTA (a Community Open Simulation, Training, and Applications framework). The first thrust focuses on developing highly-optimized graphics processing unit (GPU) implementations of common data structures and algorithms used in physical simulations, such as uniform grids, particles, octrees, and linear solvers. The second thrust consists of developing a novel differentiable physics system on top of this GPU infrastructure, based on a partial differential equation (PDE) adjoint framework. This physics system includes algorithms for fluids, solids, and their two-way coupled interactions. Given these capabilities, the project's third thrust involves building a node-based graphical user interface for building physical simulations, as well as integrations with popular AI frameworks like PyTorch. The final thrust of the project involves designing COSTA with a novel cloud-native architecture that will facilitate efficient, large-scale physics simulations on the cloud. Throughout the project, a central aim is broad community involvement, outreach, and impact. For instance, the project involves hosting public workshops on COSTA, building a community governance organization to manage COSTA development, and offering cloud computing credits and training to help onboard the greatest number of users onto the platform. Curricular materials on physics simulation, machine learning, and their combination are another result of the project, and these materials will see integration into undergraduate and graduate courses at the investigators' institutions. Further information about COSTA and its development is available at https://costaproject.github.io/. 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.