Smell of Source: Learning-Based Odor Source Localization with Molecular Communication

TL;DR

This paper compares Bayesian filtering, machine learning, physics-informed neural networks, and reinforcement learning for odor source localization, finding PINNs most accurate and RL fastest, in a molecular communication context.

Contribution

It introduces a comprehensive evaluation of multiple localization methods using synthetic data within a molecular communication framework, highlighting the superior accuracy of PINNs.

Findings

PINNs achieve the lowest localization error of 0.89e-6 m.

ML inversion has an error of 1.48e-6 m.

Reinforcement learning offers faster inference at 0.147 s.

Abstract

Odor source localization is a fundamental challenge in molecular communication, environmental monitoring, disaster response, industrial safety, and robotics. In this study, we investigate three major approaches: Bayesian filtering, machine learning (ML) models, and physics-informed neural networks (PINNs) with the aim of odor source localization in a single-source, single-molecule case. By considering the source-sensor architecture as a transmitter-receiver model we explore source localization under the scope of molecular communication. Synthetic datasets are generated using a 2D advection-diffusion PDE solver to evaluate each method under varying conditions, including sensor noise and sparse measurements. Our experiments demonstrate that \textbf{Physics-Informed Neural Networks (PINNs)} achieve the lowest localization error of \(\mathbf{0.89 \times 10^{-6}}\) m, outperforming \textbf{machine learning (ML) inversion} (\(\mathbf{1.48 \times 10^{-6}}\) m) and \textbf{Kalman filtering} (\(\mathbf{1.62 \times 10^{-6}}\) m). The \textbf{reinforcement learning (RL)} approach, while achieving a localization error of \(\mathbf{3.01 \times 10^{-6}}\) m, offers an inference time of \(\mathbf{0.147}\) s, highlighting the trade-off between accuracy and computational efficiency among different methodologies.