Fluid Intelligence: Physics‐Informed Neural Networks as CFD Surrogates for Indoor Airflow and Thermal Comfort

Abstract

Predicting indoor airflow is central to thermal comfort, but the computational fluid dynamics (CFD) that does it well is too costly to run inside an architectural design workflow. This study asks whether a deep-learning surrogate can stand in, and which kind. Three architectures are compared on one shared problem: a data-driven artificial neural network (ANN), a physics-informed neural network (PINN), and a graph neural network (GNN), all trained as surrogates for ?–? RANS CFD of a ventilated room. The comparison spans four room geometries, three training-set sizes, and six evaluation splits that separate interpolation from extrapolation along the velocity and thermal axes, with an ISO 7730 layer carrying the predicted fields through to thermal-comfort measures. Inside the sampled range the three models perform alike, with the GNN and PINN slightly ahead. Outside it they diverge: the PINN holds its temperature accuracy where the ANN and GNN fail, wins nearly every extrapolation case, and reaches better-than-full-data accuracy on temperature from a fraction of the training cases. The comfort predictions follow the field accuracy, leaving the PINN the most reliable. Physics-informed training earns its place when extrapolation and limited data matter, the conditions early-stage design most often meets.