Journal of Korean Society of Agricultural Engineers ISSN:1738-3692(Print) 2093-7709(Online) 한국농공학회논문집

This study was conducted to address the computational limitations of conventional finite element-based Computational Fluid Dynamics (CFD) analysis, which, despite its high accuracy, faces exponential increases in mesh density and computation time as facility size expands and boundary conditions become more complex. To overcome these challenges, a Physics-Informed Neural Network (PINNs) was applied to predict airflow inside a naturally ventilated greenhouse (7.0 m in width, 4.0 m in height, and 24.0 m in length) located at Kyungpook National University Farm. The PINNs models were constructed separately for the front-view and top-view domains. The front-view model considered the roof slope and utilized a deep neural network with four hidden layers (128–128–128–64 nodes), while the top-view model was simplified into a rectangular geometry (7.0 m × 24.0 m) with three hidden layers (64 nodes each). Both models used 2,000 collocation points and employed the Adam optimizer with an adaptive learning rate scheduler to improve computational efficiency. In addition, the Navier–Stokes equations and boundary conditions were incorporated to predict the velocity and pressure distributions. When compared with the conventional finite element-based CFD results, it was confirmed that the PINNs could reproduce physically consistent flow characteristics without mesh generation. These findings demonstrate that PINNs effectively mitigates the limitations of CFD and can serve as a foundational technology for future three-dimensional smart-farm simulations and internal environmental control studies.