Latest revision as of 16:06, 25 November 2022

Summary

In this work, we present a multi-fidelity machine learning surrogate model, which predicts comfort-related flow parameters in a ventilated room with a heated floor. The model uses coarseand fine-grid CFD data obtained using LES turbulence models. The dataset is created by changing the width aspect ratio of the rooms, inlet flow velocity, and temperature of the hot floor. The surrogate model takes the values of temperature and velocity magnitude at four different cavity locations as inputs. These probes are located such that they could be replaced by actual sensor readings in a practical case. The model's output is a set of comfort-related flow parameters. We test two multi-fidelity approaches based on Gaussian process regression (GPR), among them GPR with linear correction (LC GPR), and multi-fidelity GPR (MF GPR) or cokriging. The computational cost and accuracy of these approaches are compared with GPRs based on single-fidelity data. All of the tested multi-fidelity approaches successfully reduce the computational cost of dataset generation compared to high-fidelity GPR while maintaining the required level of accuracy. The co-kriging approach demonstrates the best trade-off between computational cost and accuracy.

Abstract

Full Paper