A neural network based finite element grid generation algorithm for the simulation of light propagation Original Research Article

Calculating the transfer of light in three-dimensional virtual environments is an inevitable feature of modern computer graphics systems. Usually it is accomplished by a finite element (FE) method where the three-dimensional surfaces of the scene geometry are cut into several sub-surfaces—a surface meshing is generated. We present a new way of creating this surface mesh: the internal structures of two neural networks trained by sample surface points and sample light rays are interpreted as such a mesh on which the finite element method is executed. The presented approach avoids several drawbacks arising with classical methods. The “neural meshing” outperforms standard techniques in terms of memory requirements and accuracy. Additionally this work presents several novel ideas of interpreting a neural network skeleton in terms of a virtual three-dimensional geometry and as a representation of light energy propagating through a three-dimensional polygon-based scene.