PREDICTING AEROSOL DISEASE TRANSMISSION RISK WITH REYNOLDS-AVERAGED NAVIER-STOKES (RANS) AND DEEP LEARNING

Abstract

The COVID-19 pandemic exposed the limitations of conventional aerosol risk-assessment tools such as ventilation-rate calculations, contact tracing, and oversimplified dispersion models in capturing the coupled airflow–particle dynamics that govern airborne disease transmission indoors. This paper develops an integrated Reynolds-Averaged Navier–Stokes (RANS) computational fluid dynamics (CFD) and deep learning framework for predicting aerosol-mediated transmission risk in indoor environments such as classrooms, hospital wards, and public-transport cabins. We present the governing continuity, momentum, and turbulence-closure equations for incompressible, particle-laden turbulent flow, a near-wall inflation-layer meshing strategy around indoor obstacles and occupants, and a Normalized Particle Concentration (NPC) risk metric. A convolutional neural network (CNN) surrogate is formulated to learn the mapping from scenario parameters to the RANS-simulated NPC field, and we benchmark small- and large-capacity network variants against one another across a combined suite of ventilation-velocity test cases. Quantitatively, a five-level grid-independence study spanning  to  cells confirmed mesh-independent solution behaviour, with the monitored probe quantity varying by less than  beyond the coarsest grid and by only  between the two finest meshes. A 420-scenario parametric library (  training /  validation /  test scenarios) was used to train and directly compare a small surrogate network (2,336 parameters) against a large surrogate network (  parameters, a -fold increase in capacity): on the held-out test set, the large network achieved  and RMSE =  NPC units, versus  and RMSE =  NPC units for the small network, despite converging in fewer training epochs (935 vs. 2,915). Across a combined suite of five ventilation-velocity test cases spanning , the large network reduced the mean relative L2 field error from  (small network) to  (large network), a relative reduction of roughly one-third at an inference cost of a fraction of a millisecond per scenario, several orders of magnitude faster than the parent numerical solve. We discuss the accuracy–capacity–generalization trade-offs of scaling to a full three-dimensional training library and the framework's translational value for ventilation design and public-health decision support.