Enhancing the accuracy of physics-informed neural networks for indoor airflow simulation with experimental data and Reynolds-averaged Navier–Stokes turbulence model

Physics-informed neural network (PINN) has aroused broad interest among fluid simulation researchers in recent years, representing a novel paradigm in this area where governing differential equations are encoded to provide a hybrid physics-based and data-driven deep learning framework. However, the lack of enough validations on more complex flow problems has restricted further development and application of PINN. Our research applies the PINN to simulate a two-dimensional indoor turbulent airflow case to address the issue. Although it is still quite challenging for the PINN to reach an ideal accuracy for the problem through a single purely physics-driven training, our research finds that the PINN prediction accuracy can be significantly improved by exploiting its ability to assimilate high-fidelity data during training, by which the prediction accuracy of PINN is enhanced by 53.2% for pressure, 34.6% for horizontal velocity, and 40.4% for vertical velocity, respectively. Meanwhile, the influence of data points number is also studied, which suggests a balance between prediction accuracy and data acquisition cost can be reached. Last but not least, applying Reynolds-averaged Navier–Stokes (RANS) equations and turbulence model has also been proved to improve prediction accuracy remarkably. After embedding the standard k–ε model to the PINN, the prediction accuracy was enhanced by 82.9% for pressure, 59.4% for horizontal velocity, and 70.5% for vertical velocity, respectively. These results suggest a promising step toward applications of PINN to more complex flow configurations.