Voxel-free neural volume reconstruction technique for volumetric flame reconstructions

Abstract

Volumetric reconstructions of transparent or translucent mediums are critical for various applications. For instance, successful reconstructions of transient and turbulent flames will assist in understanding the complex combustion mechanisms during combustion and advanced burner design. The most common method for volumetric flame combustion diagnosis is the tomographic reconstruction technique. Originating from computed tomography for medical diagnosis purposes, computed tomography of chemiluminescence (CTC) is a volumetric flame diagnostic method that utilizes two-dimensional projections of flame under limited viewing angles to reconstruct three-dimensional information of the combustion field. Typical flame reconstructions use discrete volumetric voxels to represent the flame luminosities at different spatial locations. However, this approach increases the computation costs in both weight matrix calculations and tomographic iterations. This investigation proposes a neural volume reconstruction technique (NVRT) that uses a neural network to represent the continuous flame luminosity implicitly. Besides, this investigation adopts the differentiable volume rendering (DVR) technique to train the network based on two-dimensional flame projections and does not require 3D supervision. We use both simulated flames and experimental flames to verify the capability of the proposed NVRT method against the traditional algebraic reconstruction technique (ART). Results show that the NVRT method is superior to the ART method in reconstruction fidelity, resistance to noise, and computational cost (especially RAM usage) for flame reconstructions.