Utilizing physics-informed neural networks to advance daylighting simulations in buildings

This paper introduces the novel application of Physics-Informed Neural Networks (PINNs) within the built environment. By leveraging the principles of the Radiative Transfer Equation (RTE) and the diffusion equation, the research presents an approach to integrating physical laws with machine learning (ML) for the precise and instant prediction of daylighting performance in buildings. The architecture of both PINNs model was significantly enhanced through Bayesian Optimization executed on a High-Performance Computing (HPC) environment, which reduced the computational time from an estimated 128 h to approximately 16 h per model. This process facilitated the efficient execution of 5000 optimization iterations for each model. For the RTE-based PINN, the best Mean Square Error (MSE) and Mean Absolute Error (MAE) were achieved in iterations 2416 and 2930, respectively. As for the Diffusion-based PINN, the best MSE and MAE were achieved in iterations 4384 and 1300, respectively. The final PINNs models showed high accuracy; the RTE-based model's MAE of approximately 0.70 and an MSE of 1.59, alongside the diffusion-based model's lower MAE of 0.55, although with a higher MSE of 17.36. Moreover, the paper presents an approach of integration of PINNs into 3D modeling environments which facilitates access to PINNs by the non-technical user.