Vol.25, No.2, 2025, pp. 310–318
https://doi.org/10.69644/ivk-2025-02-0310

This study explores the application of Physics-Informed Neural Networks (PINN) for predicting the pressure coefficient distribution on the windward facade of a building. The key feature of this approach is the use of a combined loss function that simultaneously accounts for experimental data and physical constraints derived from Bernoulli's principle. This integration enables PINN to combine physical rigor with high accuracy in reproducing experimental data. The overall loss function of the PINN model consists of two components: a mathematical loss function that minimises deviations of model predictions from experimental data, and a physics-based loss function, derived from Bernoulli's principle for the flow of an ideal fluid and gas, which ensures alignment with physical laws. A PINN model is developed and trained, demonstrating adaptability to various conditions, including wind direction angles and the weight of the physical loss function. The analysis reveals that the weight of the physical loss function significantly influences the prioritisation of predictions: at lower weights, the model prioritises experimental data, whereas at higher weights, it aligns more closely with physical calculations. Instances of prediction anomalies are observed, where the results deviate significantly from both data sources, highlighting the need for careful monitoring of the model's outputs. Future research directions include incorporating more advanced physical laws, such as turbulence and friction models, to improve prediction accuracy. Results demonstrate the potential of PINN for applications in building aerodynamics, effectively integrating theoretical calculations and empirical data for precise modelling.

Keywords: • Physics-Informed Neural Networks (PINN) • building aerodynamics • Bernoulli's principle • airflow modelling • combined approach • structural design • wind load