Three-dimensional reconstruction of smoke flow fields using deep learning and volumetric tomographic absorption spectroscopy

In fields such as aerospace, engine testing, mining, and atmospheric monitoring applications, light-absorbing flow phenomena like combustion soot and mine dust are prevalent. Investigating the three-dimensional (3D) structural information of these light-absorbing flows holds significant theoretical and practical value for understanding their combustion characteristics and environmental impact. This paper proposes the Smoke Volumetric Absorption (SMOVA) technique, which is a method for 3D reconstruction of smoke flow fields based on deep learning and volumetric tomographic absorption spectroscopy. By replacing traditional explicit voxel reconstruction with implicit neural representation, our approach substantially enhances reconstruction efficiency and flexibility. Simultaneously, we introduce multi-resolution hash encoding to substitute the conventional positional encoding in Neural Radiance Fields (NeRF), effectively compressing model size while accelerating both training and inference. We reconstructed smoke flow fields using SMOVA and benchmarked its performance against the Algebraic Reconstruction Technique (ART) and NeRF methods. Results demonstrate our method's superiority over ART and NeRF in reconstruction resolution and operational efficiency.