Enhancing precision of deformation in real-time interactive simulation using machine learning

Abstract

Real-time interactive simulation including deformable object has been researched for various engineering applications in the field of medical simulations or computer graphics for instance. Finite element method is a representative method to simulate the behavior of the deformable model accurately. Improvement of accuracy and quality for the simulation is expected with the model composed of dense elements, however, this results in greatly increased computational load and therefore appropriate compromise on the number of elements is required for real-time simulation. Several efficient computational methods, although, using the model or the subspace with smaller DOFs have been proposed, dynamic simulation of model with arbitrary constitutive law including accurate local deformation induced by interaction have not been considered. We, in this paper, propose a method to enhance precision for deformation of dynamic real-time interactive simulation with low-resolution model, through machine learning using a high-resolution model. High-resolution nodal displacements corresponding to low-resolution nodal displacements and velocities are stored in dataset, by applying a series of training external forces to both resolution model with arbitrary constitutive law. Fully connected network is designed, and learns the function to produce the deformation with improved precision from the motion of low-resolution model. Higher resolution deformation is visually rendered with small additional computation time, by applying the learned function to the low-resolution model with less computational load.