Physical Reservoir Computing with Dendritic Polymer Structures – Understanding Internal Dynamics, Scale-up and Integration in All-Solid-Systems

Physical Reservoir Computing is an emerging field of research as such recurrent neural networks offer power-efficient real-time classification on edge devices thereby utilizing the inherent temporal dynamics of the underlying physical, chemical or biological system. In particular, the use of organic mixed ionic-electronic conductors enables a close connection to biological systems due to the strong coupling between ionic and electronic conduction as well as the possibility to chemically modify the polymer to interact with bio-molecules. Previously, we have shown a first realization of a physical reservoir composed of polymer dendrites integrated onto a flexible substrate for on-chip heartbeat classification. However, the these systems was limited due to an insufficient complexity and the use of an external delayed feedback line (single-node reservoir), requiring significant data pre-processing and increasing the power consumption.<br/>In this contribution I will discuss the internal dynamics of such dendritic reservoirs and how we tuned them by external parameters such as the ion conductivity or ion mobility. Furthermore, I will demonstrate the integration of such reservoirs into a solid-state electrolyte, which allows us to scale-up the network to a large number of nodes covering the bandwidth relevant for bio-signal processing. The scale-up results in an improvement of the classification accuracy up to 95% without any silicon-based component. Moreover, the scale-up study shows us that the classification accuracy and power consumption can be decoupled in sparse network configurations, enabling us to achieve such a high accuracy with a power consumption of only 100nW. This integration strategy of sparse networks into a solid-state electrolyte opens up new perspectives to further increase the efficiency of intelligent edge devices based on organic semiconductors.