Neural network-based 3D reconstruction of temperature and velocity for turbulent flames from 2D measurements

Abstract

Three-dimensional (3D) high-resolution data of temperature and velocity are crucial for achieving a fundamental understanding of turbulent flames. However, existing combustion diagnostics are predominantly limited to the measurements at a point, along a line, or in a two-dimensional (2D) plane. In the present work, for the first time, the potential of neural networks for 3D reconstruction of turbulent combustion based on 2D measurements is explored, aiming to reconstruct both 3D temperature and velocity fields using data from a limited number of 2D planes. First, a novel translation approach incorporating two vector-quantized variational autoencoders (VQ-VAE) and a diffusion transformer model is developed for 3D temperature reconstruction based on 2D temperature distributions. Then, a wavenumber-based physics-informed neural networks (WN-PINNs) framework is established to derive the 3D velocity fields constrained by the momentum equation using the reconstructed 3D temperature and the 2D velocity measurements. The performance of the proposed neural networks is evaluated on two different configurations of turbulent flames, including freely propagating planar premixed combustion and swirling premixed combustion. The reconstructed temperature and velocity are compared with the high-fidelity direct numerical simulation (DNS) data both qualitatively and quantitatively. This study highlights the great potential of machine learning methods for the 3D reconstruction of turbulent flame fields, and provides new insights for the development of complementary tools for conventional diagnostic techniques to alleviate the challenges of 3D measurements in combustion research.

Novelty and significance

In the present work, the feasibility of using neural networks for 3D reconstruction of turbulent combustion from a limited number of 2D planes has been explored. A novel translation approach with transfer learning and a wavenumber-based physics- informed neural network framework have been established for 3D reconstruction of both temperature and velocity fields, which is the first of its kind. The proposed neural networks have demonstrated the capability to recover the flow and flame structures, with good agreement compared to high-fidelity direct numerical simulation data. The study highlight the potential of neural networks in bridging the gap between 2D measurements and 3D reconstructions for both scalar and velocity fields, offering new insights in the development of complementary tools for traditional combustion diagnostics.