Reconstruction and prediction of ventilation airflow field in the working face with sparse data using physics-informed neural networks

Intelligent mine ventilation is critical for smart mining, focusing on real-time prediction of airflow speed and pressure in ventilation systems. However, challenges such as complex underground environments, limited sensor data, and high Reynolds number turbulence hinder accurate airflow field prediction. This study proposes a physics-informed neural network (PINN) framework incorporating eddy viscosity coefficient νt turbulence modeling, combined with wind flow field data generated by computational fluid dynamics software Fluent for the working face. The model reconstructs the flow field under known conditions without relying on mine goaf flow field data. The inlet air velocity U is embedded into the model input for extrapolation to untrained conditions. Residual-based adaptive refinement with distribution (RAR-D) effectively optimizes the distribution of training points, improving prediction accuracy in complex flow regions. Experimental results show that RAR-D reduces the absolute error in the upper corner region by 29.4% compared to fixed sampling. In turbulence modeling, the Reynolds-averaged Navier–Stokes (RANS) model incorporating eddy viscosity coefficient νt demonstrates superior accuracy compared to the original Navier–Stokes equations and the RANS k−ω turbulence model. The optimization of physical constraints improves the model's extrapolation ability, reducing prediction error by more than 50% compared to purely data-driven approaches, thereby validating the effectiveness of the PINN method in intelligent mine ventilation.