Neural Network based Distance Estimation for Branched Molecular Communication Systems

TL;DR

This work tackles distance estimation in branched molecular communication channels by using an adapted Sliding Bidirectional Recurrent Neural Network (SBRNN) that ingests aggregated Rx observations from the Pogona macroscopic MC simulator to infer per-branch distances $d_{Tx,k-Rx}$. It benchmarks the approach against an analytical maximum-likelihood baseline and demonstrates robust performance in two-source topologies, with some degradation as the number of sources increases to four. The results illustrate the viability of data-driven channel parameter estimation in IoBNT-relevant MC systems and highlight directions for incorporating diffusion effects and exploring broader topologies. Overall, the study advances parameter estimation in MC by offering a practical, parameter-light method capable of handling multi-source branched environments.

Abstract

Molecular Communications (MC) is an emerging research paradigm that utilizes molecules to transmit information, with promising applications in biomedicine such as targeted drug delivery or tumor detection. It is also envisioned as a key enabler of the Internet of BioNanoThings (IoBNT). In this paper, we propose algorithms based on Recurrent Neural Networks (RNN) for the estimation of communication channel parameters in MC systems. We focus on a simple branched topology, simulating the molecule movement with a macroscopic MC simulator. The Deep Learning architectures proposed for distance estimation demonstrate strong performance within these branched environments, highlighting their potential for future MC applications.

Figures (8)

• Figure 1: Schematic of the topology of two Txs and one Rx in a branched tube environment. $\mathrm{Q_{Tx_1}}$ and $\mathrm{Q_{Tx_2}}$ are the flow in each branch, and $\mathrm{Q_{Rx}}$ is the flow in the Rx. ${\mathrm{d_{Tx_1-Rx}}}$ is the distance between Tx1 and Rx, and ${\mathrm{d_{Tx_2-Rx}}}$ is the distance between Tx2 and Rx.
• Figure 2: System model of the proposed distance estimation approach. The molecules which move from the $\mathrm{N}$ sources ($T_{\mathrm{x_1}}$ to $T_{\mathrm{x_n}}$), are received by a single Rx. This is done with the Pogona simulator. The aggregated time-series signal extracted from the Rx is processed by the Neural Network (NN), which outputs the estimated distances between Tx and Rx for each branch.
• Figure 3: Plot of the molecule release at each Tx and the number of molecules at Rx, both for the Pogona simulation and the analytical model, for $d_{\mathrm{Tx,1-Rx}} = 12\ cm$ and $d_{\mathrm{Tx,2-Rx}} = 24\ cm$ .
• Figure 4: Architecture of the SBRNN suited for distance estimation, with the number of layers, the number of neurons $n$ and activation function for each layer, and an output of N Txs.
• Figure 5: Schematic of the neural network algorithm and the sliding window system for the estimation of distance between each Tx and the Rx, for two Txs.