Improved methods for local airflow field simulation around pig bodies based on PINN

Computational fluid dynamics (CFD) has been widely used to simulate airflow in livestock and poultry housing. It is often required for the simultaneous high-precision simulation of both whole-space airflow and localized microenvironments. Geometric model simplification is also necessary in structurally complex, large houses. However, the pigs are highly sensitive to their immediate surroundings, which can directly determine their thermal comfort and health status. In this study, a mesh-free and data-driven approach—Physics-Informed Neural Networks (PINN)—was introduced to investigate different samplings and the representation of three-dimensional gas flow field reconstruction in a swine barn scenario. Experiments were conducted at the laboratory of China Agricultural University in November 2024. The PINN was employed in both two-dimensional (2D) and three-dimensional (3D) configurations. Experimental and CFD models were validated as benchmarks, while experimental and CFD simulation data served as comparative datasets. The root mean square error (RMSE) between the predicted values from PINN and the measurement was used as the evaluation metric for prediction error. The 2D experiments were divided into 24 groups, corresponding to four sampling methods—Random Sampling (RS), Latin Hypercube Sampling (LHS), Velocity Magnitude Importance Sampling (MIS), and Velocity Gradient Importance Sampling (GIS)—each applied at six sampling accuracy levels (0.18%, 0.36%, 0.72%, 1.08%, 1.8%, and 3.6%) to evaluate velocity field reconstruction. The RS method was sampled randomly, while the LHS method employed quasi-uniform sampling with the Latin Hypercube approach. In contrast, both the MIS and GIS methods were identified as the feature sampling techniques, with their sampling criteria using velocity magnitude and velocity gradient, respectively. The optimal K-value for proximity was determined via 5-fold cross-validation. A one-way ANOVA was performed on the sampling accuracy to assess significant differences after reconstruction. Each experimental group was subjected to multiple repeated trials. The 3D experiment focused on a single pig model in a unidirectional airflow environment. The performance of 3D PINN was explored to avoid the complexity of highly turbulent flows. The velocity fields were reconstructed in a swine barn setting. A comparison was made of the CFD simulations under the same conditions. There was a discrepancy between the PINN predictions and the measurement after graphic illustration. A quantitative analysis—including RMSE and velocity deviation percentage at the vortex region near the rear of the pig body—was conducted to validate the 3D PINN in such scenarios. The results revealed that: 1) A significant negative correlation was observed between sampling accuracy and reconstruction error (P < 0.05). The GIS reduced the prediction error by 30%, 50%, and 21%, respectively, at lower accuracy levels, compared with the rest, respectively. Conversely, the prediction error of the LHS was 2 to 3 times higher than that of the rest under higher accuracy conditions. 2) The 3D PINN was achieved in a prediction error of only 12.3% under experimental conditions, indicating better reconstruction. Both the training samples and the sample size significantly influenced the training effectiveness of the PINN, indicating strong applicability to reconstruct the airflow fields in swine barn environments. The PINN eliminated the dependency of conventional CFD approaches on the mesh generation and boundary conditions. Turbulent information was captured in high Reynolds number scenarios. The high accuracy of the reconstruction can be required in precision agriculture.