Coupling Neural Networks with classical numerical solvers for interactive simulation – CosimAI

Because of its low cost, interactive simulations are of great interest in a wide range of domains such as education, industrial risk assessment, design, etc., usually through Virtual Reality devices. The governing equations must be solved at each time step faster than the simulation time step, i.e. in real time. Otherwise, latency will be introduced and will hamper the immersion in the simulated world. In practice, real time is limited, on a standard PC, to simulations involving a few thousands degrees of freedom (dofs). With more dofs, the simulation remains offline, as in most Finite Elements software. The objective of this project is to bring offline simulations into the interactive world. To do so, computationally expensive sub-systems, i.e. sub-systems with a large number of dofs such as deformable shells or volumes, will be replaced by trained Neural Networks (NNs). After the training phase, NNs are indeed much faster to evaluate than with traditional methods, such as the finite element method.

The project consists in performing co-simulation between neural networks and our existing nonsmooth solver for interactive contact dynamics. The NNs will provide the forces of a solid body in response to a user's action, while the nonsmooth solver will incorporate the NNs within the whole interactive simulation. The first objective focuses on overcoming technical challenges pertaining to cosimulation. This will be exemplified on a specific problem that represent a wider class of problems: the dynamic contact interaction between a beam modelled by finite-elements and another beam described by a NN. The second objective consists in producing a realistic demonstrator of a robot controlled by a Virtuose haptic device interacting with a deformable volume.