Sensitivity Analysis on a Computational Model of Olfaction Using Gaussian Process, Generalized Polynomial Chaos and Neural Networks.

The process of odor capture by arrays of chemosensory hairs is a complex phenomenon influenced by multiple factors, such as hair morphology and surrounding fluid characteristics. Understanding how these parameters affect the fluid flow and odor transport is critical for advancing biological understanding and informing the potential design of bio-inspired sensing systems. However, performing comprehensive analyses with high-fidelity computational fluid dynamics models can be computationally expensive. To address this challenge, we investigate three surrogate modeling approaches - Gaussian process regression, generalized polynomial chaos and feedforward neural networks - as efficient numerical approximations to full simulations. These surrogates are trained on a limited number of simulations and then used to predict system responses efficiently. Their predictive accuracy is compared using synthetic examples across a range of scenarios. For the biological application considered here, cross-validation is employed to identify the surrogate model with the best performance. Using these surrogate models, we implement a variance-based global sensitivity analysis method to analyze the impact of uncertainty of inputs (such as morphological variables) on flow characteristics.