Synthesizing inertial particle distribution in isotropic turbulence using neural networks

Cluster and void formation are key processes in the dynamics of particle-laden turbulence. In this study, we propose machine learning (ML) models to synthesize the number density of point particles in turbulence using the enstrophy of the carrier fluid, by extending our previous work (Oujia et al., CTR Proc., 2022). The database of 2D direct numerical simulation (DNS) of homogeneous isotropic turbulence with one-way coupled inertial point particles is used for training the model. We compare the performance of autoencoder, U-Net, GAN, and diffusion model techniques to examine the statistics of the predicted fields by comparing against the DNS data. GAN and diffusion models result in the most accurate predictions. Furthermore, we investigate the inverse problem of synthesizing enstrophy fields using particle density distribution as input, for various particle Stokes numbers. The results of this study indicate the potential use of the ML techniques in predicting turbulence features from experimental inertial particle measurements.