Applying a physics-informed neural network to an indoor airflow time-extrapolation prediction

Deep learning (DL) methods have been recently explored as alternatives to computational fluid dynamics for indoor airflow prediction. However, traditional pure data-driven DL models typically struggle with time-extrapolation predictions, exhibiting significant accuracy degradation when the testing data deviates notably from the training data. To address this limitation, a physics-informed neural network (PINN) was applied, integrating data with governing physical equations to improve the time-extrapolation prediction accuracy of indoor airflow dynamics. In this study, the PINN is trained on early-stage vortex development in a room and subsequently tasked with predicting future airflow evolution. Compared to a pure data-driven artificial neural network (ANN), the PINN achieves substantially reduced prediction errors: the absolute errors of velocity magnitude, x-velocity component, y-velocity component, and pressure are 84.2 %, 72.5 %, 77.8 %, and 98.3 % of the ANN’s errors, respectively, demonstrating significant accuracy enhancements. Furthermore, the PINN preserves vortex integrity over time, whereas the vortex structure predicted by the ANN deteriorates. Additional analysis shows that the PINN is more robust even under incomplete boundary conditions. These findings indicate that incorporating physical constraints into DL-based time-extrapolation predictions leading to more robust and physically consistent results.