A Fourier neural operator framework for indoor contaminant transport prediction, source localization, and release strength estimation

Source abstract

Indoor air quality is a key concern in building design, as occupant exposure is strongly governed by pollutant transport and removal through ventilation flows. Computational fluid dynamics (CFD) provides detailed indoor contaminant concentration fields, but its computational cost limits repeated evaluations over many pollutant source locations. In this work, we develop a Fourier Neural Operator (FNO) framework for rapid prediction and inverse localization of indoor contaminant transport, in which a soft rectangular source mask provides both geometric fidelity and differentiability. The FNO achieves a mean relative L2 field error of 0.032, predicts a full field in approximately 5.5 ms, about 15,000 times faster than the corresponding CFD simulation, and outperforms FCN, U-Net, and DeepONet baselines. Zero-shot super-resolution tests show that models trained on moderately coarse grids can be evaluated directly on finer grids with little loss of accuracy: the model trained at 192 × 64 increases the mean root-mean-square error by only 14.0% on the 384 × 128 reference grid. For inverse localization, the differentiable surrogate is combined with coarse global screening and continuous gradient-based refinement to estimate source coordinates from eight sensor observations, achieving mean localization errors of 0.103, 0.165, and 0.372 m at relative noise levels of 0, 0.02, and 0.05, the lowest among the three evaluated inverse strategies. These results demonstrate that differentiable neural-operator surrogates can support rapid field prediction and reliable inverse source localization for indoor contaminant transport