Ventilation flow reconstruction in a simplified airplane cabin model using measured mean velocities and physics-informed neural networks

Ventilation flows are investigated using experiments and field measurements. However, these methods have limitations in capturing the entire complex ventilation flow in detail owing to the limited number of data points. In this study, we propose a method for reconstructing high-resolution ventilation flows from sparse measurement data using physics-informed neural networks (PINNs). To verify the effectiveness of this technique, we used an experiment that measured the ventilation flow in detail in a reduced-scale simplified airplane cabin through particle image velocimetry (PIV). We extracted the mean velocity components and magnitudes and randomly selected different portions of measurement points from the PIV data. These values were then fed to the PINNs as observed data along with physical information, such as boundary conditions and Navier–Stokes equations, to reconstruct the mean velocity distributions. This approach accurately reconstructed the main flows, even with a few hundred observation points. However, with fewer points, the secondary flows became less clear, thus requiring additional constraints on the velocity directions at the inlets and outlets. The method also worked with only mean velocity magnitudes; however, directional constraints became critical as the observation points decreased, leading to partially misoriented flows.