Task-Adaptive Physical Reservoir Computing via Tunable Molecular Communication Dynamics

TL;DR

This paper tackles the rigidity of conventional physical reservoir computing by demonstrating that a canonical Molecular Communication channel can serve as a task-adaptive PRC. It employs a dual-simulation approach—deterministic mean-field and particle-based stochastic (Smoldyn)—and Bayesian optimization to map biophysical parameters to memory-dominant or nonlinearity-dominant computational regimes. The key findings show memory-rich parameters excel at chaotic time-series forecasting while high nonlinearity enables nonlinear data transformation, with stochastic noise mitigated by temporal averaging. The work provides a design blueprint for tunable, bioinspired wetware AI systems and outlines a path toward online, real-time reconfigurability of reservoir dynamics.

Abstract

Physical Reservoir Computing (PRC) offers an efficient paradigm for processing temporal data, yet most physical implementations are static, limiting their performance to a narrow range of tasks. In this work, we demonstrate in silico that a canonical Molecular Communication (MC) channel can function as a highly versatile and task-adaptive PRC whose computational properties are reconfigurable. Using a dual-simulation approach -- a computationally efficient deterministic mean-field model and a high-fidelity particle-based stochastic model (Smoldyn) -- we show that tuning the channel's underlying biophysical parameters, such as ligand-receptor kinetics and diffusion dynamics, allows the reservoir to be optimized for distinct classes of computation. We employ Bayesian optimization to efficiently navigate this high-dimensional parameter space, identifying discrete operational regimes. Our results reveal a clear trade-off: parameter sets rich in channel memory excel at chaotic time-series forecasting tasks (e.g., Mackey Glass), while regimes that promote strong receptor nonlinearity are superior for nonlinear data transformation. We further demonstrate that post-processing methods improve the performance of the stochastic reservoir by mitigating intrinsic molecular noise. These findings establish the MC channel not merely as a computational substrate, but as a design blueprint for tunable, bioinspired computing systems, providing a clear optimization framework for future wetware AI implementations.

Figures (3)

• Figure 1: Conceptual framework of the task-adaptive MC-PRC. The diagram illustrates the MC channel (transmitter, diffusion, receiver) with labels for the "Biophysical Control Knobs" (D, d, kon, etc.). Three equations are labeled and shown to explain the mechanism of the PRC which are analyzed further in Section 2A.
• Figure 2: a)Task-Specific Clustering of Optimal Biophysical Parameters. A parallel coordinates plot visualizing the top 10 best-performing parameter sets discovered by Bayesian optimization(200 deterministic model trials) for each of the three computational tasks. Each colored line represents a single, complete parameter set. The vertical axes correspond to the seven(also $K_D$, defined as $k_{off}/k_{on}$) biophysical parameters that were optimized. Clustering patterns emerge between the Forecasting task (blue), the Transformation task (red) and the Hybrid task (green). This visualization provides strong insight for the possibility of distinct, task-adaptive operational regimes in the MC-PRC. The "Memory Window Length" refers to a readout hyperparameter, included in the optimization, defining the number of recent reservoir states used for prediction. b)Demonstration of task-adaptability through performance comparison. The chart shows the NRMSE error for three distinct computational tasks on deterministic model when performed with parameter sets optimized for each specific task. The lowest error for each task is achieved only when using its corresponding matched parameter set, visually confirming that reconfiguring the reservoir's biophysical parameters is critical for high performance.
• Figure 3: Demonstration of task-adaptability through performance comparison on three distinct computational benchmarks wringe2025reservoir. Each plot shows the target ground truth (black), the prediction from the idealized deterministic model (blue), and the averaged prediction from three particle-based stochastic Smoldyn simulations with moving average window (red). The results for each task were generated using a unique set of biophysical parameters identified via Bayesian optimization. (a) Forecasting Task: On the memory-dominant Mackey-Glass time-series prediction with 6 symbols ahead, the deterministic achieved an NRMSE of 0.097 whereas filtered stochastic(Mov. Av. Window of 2000) achieved 0.4927. (b) Nonlinear Transformation Task: For the nonlinearity-dominant sine-to-square wave transformation, the deterministic achieved an NRMSE of 0.237 whereas filtered stochastic(Mov. Av. Window of 6000) achieved 0.3882. (c) Combination Task: The model also performs effectively on the Mackey-Glass Cubed task with 10 symbols ahead, which requires a balance of both memory and nonlinearity, the deterministic achieved an NRMSE of 0.307 whereas filtered stochastic(Mov. Av. Window of 2000) achieved 0.7428.