Supervised learning for boundary detection on particle systems

Abstract

In particle-based physics simulations, the information about which particles belong to the boundary of the system and which are considered internal is, in general, an information that is useful but hard to obtain in an efficient way. This information can be applied to generate the free surface of the fluid or to compute the surface tension, among other applications. Techniques found in the literature may present satisfactory results, but in general they are sensible to the problem scale, particle distribution and involve computationally expensive operations such as matrix inversion. The goal of this study is to present an alternative with lower computational cost at runtime by transferring some of the work to a preprocessing step using offline learning.