Figure 1: We learn fast subspace models on-the-fly to accelerate deformable simulation: (Top) Ground truth frames from a 1000-frame deformable character simulation with 165,941 tetrahedra and Arruda-Boyce material which took 48 hours (174,521 seconds). (Middle) Frames from our online integrator which were generated in only 51 minutes (3,101 seconds) -56.28× faster -by learning a fast 11-dimensional subspace model on-the-fly (Bottom) Visualization of the temporal behavior of the online integrator reveal full steps (in black), reduced steps (in gray), and basis updates (red ticks). A reduced basis is detected quickly, and the first "skip" occurs at timestep 2. Only a handful of full steps are needed, and our reduced solver computes over 98% of the the steps.
AbstractFinite element simulations of nonlinear deformable models are computationally costly, routinely taking hours or days to compute the motion of detailed meshes. Dimensional model reduction can make simulations orders of magnitude faster, but is unsuitable for general deformable body simulations because it requires expensive precomputations, and it can suppress motion that lies outside the span of a pre-specified low-rank basis. We present an online model reduction method that does not have these limitations. In lieu of precomputation, we analyze the motion of the full model as the simulation progresses, incrementally building a reduced-order nonlinear model, and detecting when our reduced model is capable of performing the next timestep. For these subspace steps, full-model computation is "skipped" and replaced with a very fast (on the order of milliseconds) reduced order step. We present algorithms for both dynamic and quasistatic simulations, and a "throttle" parameter that allows a user to trade off between faster, approximate previews and slower, more conservative results. For detailed meshes undergoing low-rank motion, we have observed speedups of over an order of magnitude with our method.