Sensitivity and uncertainty analyses for model of olfaction using Dempster-Shafer theory and neural networks.

Olfactory systems in arthropods rely on arrays of chemosensory hairs (sensilla) that interact with the surrounding fluid to capture odor cues. The fluid dynamics depend on both morphological parameters and flow conditions, and variability in these inputs introduces both stochastic and epistemic uncertainties. In this study, we introduce an integrated computational framework that combines Dempster-Shafer (DS) theory with neural-network-based surrogate modeling to quantify and analyze mixed uncertainties in a two-dimensional computational fluid dynamics (CFD) model of olfaction. Specifically, belief functions are constructed for the epistemic variables (e.g., inter-sensilla gap) associated with limited number of samples, and feedforward neural networks are trained as computationally efficient surrogates for the full CFD simulations. These surrogates enable efficient global sensitivity analysis (SA), which identifies the influential uncertain factors affecting key quantities of interest (QoIs). We find that leakiness is most sensitive to uncertainty in the sensilla array orientation (angle) and inter-sensilla gap, and mean velocity is primarily influenced by uncertainty in the array angle and Reynolds number. Guided by the SA results, output uncertainty is characterized through DS belief structures over the expectation of QoIs when the epistemic variables are influential (such as leakiness). When the uncertainty is predominantly stochastic (e.g., mean velocity), it is instead represented using empirical probability density functions.