Abstract : This study proposes an neural network framework to model foam effects found in liquid simulation without noise. The position and advection of foam particles are calculated using the existing screen projection method, and the noise problem that occurs in this process is overcome by using an neural network. An important problem in the screen projection approach is the noise generated in the projection map in the process of projecting momentum onto the discretized screen space. We efficiently solve this problem by utilizing an neural network-based denoising network. Following the selection of the foam generation area using projection map, the foam particles are generated through the inverse-transformation of the 2D space into 3D space. To this end, we solve the problem of small-sized foam dissipation, which occurs in conventional denoising networks. Furthermore, by integrating the proposed algorithm with the screen-space projection framework, it is able to maintain all the advantages of this approach. In conclusion, owing to the denoising process and clean foam effects, the proposed network can stably model the foam effects.

[paper]