A CFD-based multi-fidelity surrogate model for predicting indoor airflow parameters using sensor readings

In this study, we introduce a multi-fidelity machine learning surrogate model that predicts comfort-related flow parameters in a benchmark scenario of a ventilated room with a heated floor. The model leverages both coarse- and fine-grid CFD simulations employing a LES turbulence model. To build the dataset, we varied parameters such as the room's width aspect ratio, inlet flow velocity, and the heated floor's temperature. The surrogate model inputs temperature and velocity magnitude readings from two specific locations, chosen to simulate real sensor placements, thus enhancing its applicability in practical scenarios. The model's outputs include the average Nusselt number on the heated wall, the jet separation point, average kinetic energy, average enstrophy, and average temperature. We explore three multi-fidelity approaches: Gaussian process regression (GPR) using combined high- and low-fidelity data without differentiation, GPR with linear correction for the low-fidelity data, and multi-fidelity GPR or co-kriging. These methods are evaluated for their computational cost and accuracy against GPR models relying solely on high-fidelity or low-fidelity data. All multi-fidelity approaches effectively reduce the computational expense of dataset creation while maintaining required accuracy levels. Among them, co-kriging offers the best balance between computational cost and accuracy.