A theoretical framework for reservoir computing on networks of organic electrochemical transistors

TL;DR

The paper tackles predicting complex dynamics when governing rules are hard to learn by proposing a theoretical framework for physical reservoir computing using networks of organic electrochemical transistors (OECTs) as nonlinear units. It develops a mathematical model of OECT networks, outlines training and autonomous prediction procedures, and validates the approach by reproducing the Lorenz attractor with predictive performance comparable to standard reservoir computers for modest reservoir sizes. A key finding is the strong dependence of forecast horizon on the pinch-off voltage $V_{ ext{p}}$, with more nonlinear operation yielding better long-term predictions, while network connectivity shows limited impact. Overall, the work provides a design-and-analysis roadmap for building physical reservoir computers with OECTs and highlights device-parameter regimes that maximize predictive capability.

Abstract

Efficient and accurate prediction of physical systems is important even when the rules of those systems cannot be easily learned. Reservoir computing, a type of recurrent neural network with fixed nonlinear units, is one such prediction method and is valued for its ease of training. Organic electrochemical transistors (OECTs) are physical devices with nonlinear transient properties that can be used as the nonlinear units of a reservoir computer. We present a theoretical framework for simulating reservoir computers using OECTs as the non-linear units as a test bed for designing physical reservoir computers. We present a proof of concept demonstrating that such an implementation can accurately predict the Lorenz attractor with comparable performance to standard reservoir computer implementations. We explore the effect of operating parameters and find that the prediction performance strongly depends on the pinch-off voltage of the OECTs.

Figures (8)

• Figure 1: Implementation of a reservoir computer using a recurrent neural network.
• Figure 2: a) A conceptual illustration of an example multi-gate OECT that serves as a node in the reservoir computer, in this case with three gate electrodes that receive signals from neighboring nodes. Also shown is the circuit symbol that is used in panel c). b) The circuit diagram of the device, showing the parameters used in the theoretical OECT model. c) The circuit diagram of an example OECT RC with a single representative connection between the drain of OECT $i$ and the gate of OECT $j$, weighted with resistor $\bm{R}_{\text{w}}^{ij}$.
• Figure 3: A 3D plot of the Lorenz attractor (grey lines) and the prediction of the Lorenz attractor using an OECT RC (blue lines) of size $N=100$.
• Figure 4: A plot of the $x$, $y$, and $z$ components of the Lorenz attractor for the true time series (solid grey line), the time series predicted using the RC implementation described in Ref. pathak_using_2017 (teal dashed-dotted line), and the time series predicted using the theoretical model of an OECT RC (blue dashed line). In both cases, $N = 100$.
• Figure 5: A plot of the forecast horizon of the Lorenz attractor described in Eqs. \ref{['eq:lorenz1']}-\ref{['eq:lorenz3']} with respect to the reservoir size, $N$, for (1) an RC constructed according to Ref. pathak_using_2017 (teal dash-dotted line) and (2) an OECT RC (blue dashed line). The prediction task is repeated $10^2$ times for each reservoir size. The connected lines represent the mean forecast horizon and the error bars represent the standard deviation in the forecast horizon.