Development of metamodels for predicting aerosol dispersion in ventilated spaces

Artificial neural network (ANN) based metamodels were developed to describe the relationship between the design variables and their effects on the dispersion of aerosols in a ventilated space. A Hammersley sequence sampling (HSS) technique was employed to efficiently explore the multi-parameter design space and to build numerical simulation scenarios. A detailed computational fluid dynamics (CFD) model was applied to simulate these scenarios. The results derived from the CFD simulations were used to train and test the metamodels. Feed forward ANN’s were developed to map the relationship between the inputs and the outputs. The predictive ability of the neural network based metamodels was compared to linear and quadratic metamodels also derived from the same CFD simulation results. The ANN based metamodel performed well in predicting the independent data sets including data generated at the boundaries. Sensitivity analysis showed that particle tracking time to residence time and the location of input and output with relation to the height of the room had more impact than the other dimensionless groups on particle behavior.

Research highlights► The paper outlines a systematic method for developing predictive metamodels for particle behavior in ventilated spaces from computational fluid dynamic simulations. ► Compared to linear and quadratic metamodels, artificial neural network based metamodels predicted particle behavior more accurately. ► The study showed a way of analyzing a very complex and highly non-linear problem and obtaining meaningful insight. ► Sensitivity analysis showed that particle tracking time to residence time and the location of input and output with relation to the height of the room had more impact than the other dimensionless groups on particle behavior.