Efficient learning representation of noise-reduced foam effects with convolutional denoising networks

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Fluid Dynamics Simulation

Background:

• Screen projection methods in liquid simulation often generate noise, particularly in the projection map during momentum transfer.
• Conventional denoising networks can lead to the dissipation of small foam particles, impacting visual fidelity.

Purpose of the Study:

• To propose a neural network framework for stable and noise-free modeling of foam effects in liquid simulations.
• To address the noise problem inherent in screen projection methods for foam particle advection.
• To prevent the dissipation of small foam particles, a common issue in existing denoising techniques.

Main Methods:

• Utilizing a denoising neural network to eliminate noise in the screen projection map.
• Calculating foam particle position and advection via screen projection.
• Generating foam particles through inverse transformation from 2D to 3D space after identifying foam areas.
• Integrating the neural network algorithm with the existing screen-space projection framework.

Main Results:

• Successfully prevented noise generation in the projection map during foam simulation.
• Enabled stable modeling of foam effects by combining denoising with screen-space projection.
• Resolved the issue of small-sized foam particle dissipation, enhancing realism.
• Maintained the advantages of the screen-space projection framework.

Conclusions:

• The proposed neural network framework effectively models foam effects in liquid simulation stably and without noise.
• The integration of a denoising neural network with screen-space projection offers a robust solution for realistic foam rendering.
• This method overcomes limitations of previous approaches, leading to improved visual quality and stability in simulated foam.