Brain-inspired polymer dendrite networks for morphology-dependent computing hardware

TL;DR

This work presents brain-inspired, morphology-dependent hardware based on AC-electropolymerized PEDOT:PSS dendritic networks that operate in aqueous environments for in materio computing. It shows that network topology induces nonlinear and memory-capable dynamics through self-gating and inter-gating effects, enabling multiply-accumulate operations and spatiotemporal information processing with a single readout. The study demonstrates hardware-dependent functions and device-specific signatures arising from stochastic dendrite growth, suggesting potential for physically unclonable functions and secure neuromorphic hardware. Overall, electropolymerization offers a scalable bottom-up route to plastic, cost-effective organic neuromorphic systems that exploit topological richness for computation.

Abstract

Variability has always been a challenge to mitigate in electronics. This especially holds true for organic semiconductors, where reproducibility and long-term stability concerns hinder industrialization. By relying on a bio-inspired computing paradigm, we show that AC-electropolymerization is a powerful platform for the development of morphology-dependent computing hardware. Our findings reveal that electropolymerized polymer dendrite networks exhibit a complex relationship between structure and operation that allows them to implement nearly linear to nonlinear functions depending on the complexity of their structure. Moreover, dendritic networks can integrate a limitless number of inputs from their environment, for which their unique morphologies induce specific patterns in the dynamic encoding of the network's output. We demonstrate that this property can be used to our advantage in the context of in materio computing to discriminate between different spatiotemporal inputs. These results show how, due to its inherent stochasticity, electropolymerization is a pivotal technique for the bottom-up implementation of computationally powerful objects. We anticipate this study will help shifting the negative perception of variability in the material science community and promote the electropolymerization framework as a foundation for the development of a new generation of hardware defined by its topological richness.

Figures (9)

• Figure 1: Nonlinear behavior and autogating effect of a multiterminal dendritic OECT. a, Microscopic photo of a four-terminal dendritic system. Scale bar = 100 µm. b, Schematic of ionic distribution within the system when different potentials are applied. Relative potential is represented by the color scale, red being the point of highest potential in the system while blue is the lowest one. c, Output characteristics of the Y-shaped device pictured in a). The curve shows an asymmetric electrical behavior related to the distribution of ions in the system, as illustrated in the subset figures. When the applied voltage is negative, the output current reaches a plateau at about 0.4 V, whereas it remains linear when V > 0 and up to 0.9 V. The rectification coefficient, computed as $|I_{OUT}|_{0.9V}/ |I_{OUT}|_{-0.9V}$, is equal to 4.4. d, Breakdown of the rectification coefficients as a function of which terminals are polarized. The x-axis indicates which electrodes are addressed in each configuration (e.g.,(1·2)-3). The grounded electrode is always written to the left of the hyphen (electrodes 1 and 2 are shorted), and the terminal undergoing voltage sweep is always written to the right (electrode 3). The numbers correspond to the electrodes indicated in the inset of the figure.
• Figure 2: Inter-gating effect between two parallel dendritic OECTs. The morphology of the PEDOT:PSS fibers is of primary importance. a, Microscopic photo of the two dendrites used as OECTs, showing that the left one (grown at 25 Hz) is noticeably thicker than the right one (grown at 200 Hz). b, Equivalent circuit of the two parallel transistors. The PEDOT:PSS channel of each OECT act has the gate from the other transistor. c-d, Ionic distribution when the bulky and thin dendrites are respectively used as the gate. The color scale indicates the nodes of highest potential (red) and lowest potential (blue) in the circuit. e-f, Transfer characteristics of the thin and bulky dendrites, respectively. g-h, Output characteristics of the thin and bulky dendrites, respectively.
• Figure 3: Real-time computing, taking advantage of the non-linearities and inter-gating effect of the dendritic OECTs. a-b, Microscopic photo and schematics of the system showing the three inputs and the output. The Y-shaped device is used as the input, while the output current is read at the drain of the single dendrite ($V_{READ}$ = 100 mV). c, Output current and input voltages as a function of time. Transients were manually removed to focus on the steady-state response of the circuit. d, Ionic distribution schematizing the dedoping mechanisms in the system when the inputs are polarized. The color scale indicates the nodes of highest potential (red) and lowest potential (blue) in the circuit.
• Figure 4: In memory spatial information processing. a, Principle diagram showing the WRITE and READ operation sequence. Each color stands for a 3-bit voltage pattern. b, Microscopic photo of the dendritic network used for spatial information processing. c, Output current variation for five repetitions of Spatial Projection #1. The output current variation is computed as the difference between the output current measured after a WRITE operation and the previous REST operation (all terminals set to the ground for ten seconds). d, Output current variation for different sets of inputs receiving the temporal sequence presented in a). In the schematics of the dendritic system on the right, each letter represents the position of a single bit in the 3-bit pattern (A,B,C). e, Spider diagram showing the average current variation and standard deviation over five cycles for three different sets of inputs, as presented in d). Chronological order is shown clockwise, starting from the top of the diagram.
• Figure 5: In memory temporal information processing. a, Microscopic photo of the dendritic network used for temporal information processing, the same as in Figure 4. b, Output current variation for different temporal sequences (the 3-bit patterns ore the same, only shuffled) presented to the same input electrodes in the system. Each color represent a 3-bit pattern. Inset : schematic representation of the dendritic system, for which blue represents a HIGH logic state (physically encoded as 600 mV), and red represents a LOW logic state (physically encoded as -600 mV). c, Spider diagram showing the average current variation and standard deviation over five cycles for two different temporal sequences. Note that following the diagram clockwise does represent the chronological order for temporal projection #2, but not for #3.