Prediction of Reynolds and stokes number behavior of particle-laden turbulent flows using artificial neural networks

A hybrid machine learning framework is proposed, demonstrated, and assessed for the prediction of inertial particle dynamics in turbulent channel flows. The approach couples a feed-forward artificial neural network, trained on direct numerical simulation (DNS) and Lagrangian particle tracking (LPT) data, with a stochastic Gaussian noise model to reconstruct particle trajectories leading to the prediction of both the mean and fluctuating particle velocity fields. The model extends prior work by incorporating lagged particle velocities, as well as local particle concentrations and concentration gradients, to capture inertial memory effects and by introducing a novel wall-normal-dependent noise generation method to reproduce spatially varying turbulence intensity. The training dataset spans two Reynolds numbers (Reτ=180 and 300), three particle diameters (dP*=0.0025, 0.005, and 0.0075), and three density ratios (ρP*=2.5, 1111.0, and 2041.0), covering a broad range of Stokes numbers (St+=0.028 → 574.03). The hybrid model reproduces the correct qualitative trends for key turbulent transport features, including inertial lag, turbophoresis, and enhanced near-wall velocities for inertial particles, and recovers root-mean-square velocity fluctuations well, particularly for high Stokes number particles. Visual comparisons of particle trajectories further confirm the model's ability to capture large-scale dynamics across particle-laden flow regimes. Validation against DNS-LPT shows good agreement for both first- and second-order statistics. A held-out test at a Reynolds number absent from the training data confirms that the model interpolates across the Reynolds-number range. The framework performs consistently across the range of particle and flow parameters studied and provides a computationally efficient alternative to full-resolution simulations for modeling particle-laden turbulent flows.