Abstract: Although numerical simulation methods have developed rapidly in solving turbulent processes in fluid dynamics, there are still challenges in accurately modeling and computing speed when dealing with complex geometric shapes and flow processes. In response to the current high computational cost issues in Computational Fluid Dynamics (CFD), this paper combines traditional turbulence numerical simulation techniques with the machine learning. Taking the classic Sandia Flame D combustion model as an example, by introducing deep learning algorithms of physical information, a Physical Information Neural Network (PINN) architecture is established. The physical information that conforms to the rules is embedded into the neural network, so that parameter flow field reconstruction can be achieved with small samples. On the plane dimension, the reconstruction results of PINN and data-driven methods were compared and analyzed with the simulation results of CFD software. The PINN method can obtain the reconstruction results of data-driven methods in large sample situations when the training set size is less than half of the total number of sample points. The L₂ relative errors of the reconstructed axial and radial velocities and temperatures of the combustion process at t=1 s were 0.187%, 1.194%, and 0.071%, respectively. Moreover, when the training set accounts for 55%, 70%, and 82% of the total number of sample points, the PINN method has smaller errors than data-driven methods. In terms of time dimension, the axial velocity cloud maps at t=0.3 s, 0.5 s, and 1 s were successfully reconstructed, proving that the PINN method can reconstruct the physical field distribution cloud map at any time within the sampling time range of the geometric model.