Reservoir Computing-Based Detection for Molecular Communications

TL;DR

The paper addresses the difficulty of reliable detection in diffusion-based molecular communications under severe inter-symbol interference and mobility, where channel state information is hard to obtain. It proposes a reservoir computing detector using a fixed nonlinear reservoir (echo state network) and a trainable linear readout to exploit temporal dependencies with low complexity, including two strategies: standard RC and RC-ISI with ROC-optimized thresholds. The authors benchmark against MAP, EMA, and several deep learning detectors in a realistic 3D Smoldyn mobile MC setting, showing RC-ISI maintains superior robustness across high ISI, with ultra-low inference latency (~1 μs per symbol) and a tiny parameter footprint (≈401 trainable weights). The approach demonstrates that reservoir computing can achieve competitive or superior performance with much lower training and computational costs, making it attractive for micro/nano-scale receivers and potential wetware implementations.

Abstract

Diffusion-based Molecular Communication (MC) is inherently challenged by severe inter-symbol interference (ISI). This is significantly amplified in mobile scenarios, where the channel impulse response (CIR) becomes time-varying and stochastic. Obtaining accurate Channel State Information (CSI) for traditional model-based detection is intractable in such dynamic environments. While deep learning (DL) offers adaptability, its complexity is unsuitable for resource-constrained micro/nanodevices. This paper proposes a low-complexity Reservoir Computing (RC) based detector. The RC architecture utilizes a fixed, recurrent non-linear reservoir to project the time-varying received signal into a high-dimensional state space. This effectively transforms the complex temporal detection problem into a simple linear classification task, capturing ISI dynamics without explicit channel modeling or complex retraining. Evaluated in a realistic 3D mobile MC simulation environment (Smoldyn), our RC detector significantly outperforms classical detectors and achieves superior performance compared to complex ML methods (LSTM, CNN, MLP) under severe ISI. Importantly, RC achieves this with significantly fewer trainable parameters (e.g., 300 vs. up to 264k for MLP) and ultra-low latency inference (approx. 1 $μ$s per symbol).

Figures (5)

• Figure 1: Diffusion-based molecular CIR and accumulated ISI from consecutive symbols. Non-ML detection in diffusion-based MC for different $T_\mathrm{b}$. Blue: received signal; orange: adaptive (EMA) threshold; dotted: fixed threshold. As $T_\mathrm{b}$ decreases, ISI raises the baseline, and the adaptive detection better tracks it than the fixed threshold.
• Figure 2: Overview of the proposed RC-ISI detection architecture. The mobile MC environment is simulated (1), and the received signal (receptor occupancy) is normalized and fed into the fixed recurrent reservoir (2). The high-dimensional reservoir states are then mapped by the trained readout layer, followed by ROC-based thresholding (3) to produce the final bit decision.
• Figure 3: ROC curve for RC-ISI at $T_\mathrm{b}{=}100$s. The high AUC (0.954) indicates strong classification performance and separability between bit $1$ and bit $0$.
• Figure 4: Accuracy comparison versus symbol time ($T_\mathrm{b}$) for all evaluated detectors. RC-ISI maintains the highest robustness as $T_\mathrm{b}$ decreases (ISI increases), significantly outperforming classical methods and complex ML models under severe ISI.
• Figure 5: Inference latency ($\mu$s, log scale) of detectors across different symbol durations. RC and RC-ISI achieve the lowest latency among ML methods, comparable to simple MLP implementations and significantly faster than LSTM/CNN.