Reconfigurable neuromorphic memristor network for ultralow-power smart textile electronics

As conventional complementary metal-oxide-semiconductor integrated circuits approach physical limits, in-memory computing has emerged as a low-power, high-efficiency alternative inspired by brain architectures such as ANNs and SNNs. Core elements, electronic synapses and neurons, typically require different materials and fabrication processes, complicating heterogeneous integration and limiting density. Performance mismatch between discrete synaptic devices and neurons further impedes cooperative neural networks, motivating reconfigurable memristors that can serve as both synapse and neuron. Electronic textiles integrating neuromorphic memristors are promising for wearable systems requiring low energy. Biological neurons operate at ~1–100 pJ/spike, guiding targets for ultralow-power fiber-based devices. Typical artificial neurons rely on multi-component circuits (memristor, capacitor, resistor), adding complexity and power overhead, highlighting a need for simplified, low-power neuron functions. Herein, the authors fabricate a reconfigurable textile memristor network using an Ag/MoS2/HfAlOx/CNT heterostructure that exhibits both nonvolatile resistive switching (synapse) and volatile threshold switching (neuron) by regulating conductive filaments. The nonvolatile mode provides multi-level synaptic weights with long-term storage; the volatile mode enables integrate-and-fire spiking from prior neuron inputs. A single device realizes neuron function with ultralow firing energy of 1.9 fJ/spike, three orders of magnitude below biological neurons. Integrating reconfigurable synapses and neurons with a heating resistor demonstrates an automatic, intelligent heating textile, pointing toward next-generation, in-memory neuromorphic textiles.

Prior neuromorphic systems often implement synapses and neurons with different materials and device structures, causing fabrication incompatibilities and performance mismatches that hinder dense, cooperative networks. Reconfigurable devices have been proposed to unify logic and neuromorphic functions, but textile-oriented, fiber-shaped implementations remain limited. In wearable electronics, electronic textiles with display, sensing, energy harvesting, and storage have advanced, yet neuromorphic processing integrated seamlessly into textiles with ultralow power remains an unmet need. Conventional artificial neurons typically require multi-component circuits (memristor-capacitor-resistor), increasing complexity and energy use. Biological neurons operate at picojoule-per-spike energy, setting a benchmark that most artificial devices (oxide, 2D materials, others) exceed by orders of magnitude. This work addresses these gaps by using a single reconfigurable fiber memristor capable of both synaptic and neuron behaviors with femtojoule-level energy.

Device architecture: Fiber-based memristor units comprise an Ag top electrode, MoS2 nanosheet layer, HfAlOx switching layer (~20 nm), and a CNT fiber bottom electrode, assembled into a 3D woven textile network with interwoven top and bottom layers acting as synapses and neurons, respectively. Fabrication: Ag fibers were sequentially cleaned in acetone, isopropanol, and deionized water (5 min each). MoS2 nanosheets were deposited on Ag fibers via electrophoretic deposition in ethanol. A 20 nm HfAlOx layer was deposited by atomic layer deposition (ALD) at 130 °C using a cycle sequence of trimethylaluminum (TMA), H2O, tetrakis(ethylmethylamino)hafnium (TEMAH), and H2O, with Ar as carrier gas. The memristor was formed by interlacing the coated Ag fiber with a CNT fiber. CNT fibers were spun from CVD-grown CNT arrays (1250 °C) using ethanol/acetone as carbon source and ferrocene as catalyst; these CNT fibers also served directly as heating resistors. Characterization: Morphology and structure were examined via field-emission SEM (ZEISS SIGMA HD) and cross-sectional TEM (Talos-F200). Electrical behavior was measured using an Agilent B1500A (DC sweeps) and pulse tests via Semiconductor Pulse Generator Unit and waveform generator/fast measurement unit in air. Conductive AFM (XE-100) provided current maps to study filament formation under varying scan voltages. Operating modes: Nonvolatile resistive switching (synaptic mode) and volatile threshold switching (neuronal mode) were selected by setting compliance currents; higher compliance (e.g., 100 µA, 1 mA) formed stronger filaments for nonvolatile switching, while lower compliance (e.g., 10 µA, 1 µA, 100 nA) produced weaker, volatile filaments for neuron-like threshold switching. Synapse emulation used pulse trains to program/erase multi-level conductance and measure EPSC, PPF/PPD, LTP/LTD, endurance, and retention (100 mV read). Neuron emulation applied consecutive pulses (1 ms width, 0.5–1.5 V amplitude) with a firing threshold set at 1 pA to demonstrate integrate-and-fire behavior and quantify firing energy E = V × Ifiring × t. System integration: A smart textile prototype integrated reconfigurable synapses and neurons with CNT fiber resistors as heaters. The synapse modulated weights and thus voltage division to the neuron, which translated spiking outputs into heating operation timing. A control cycle used 60 s periods for neuromorphic unit initialization and a 3000 s heating period for the resistor-based thermal actuation.

• A reconfigurable textile memristor (Ag/MoS2/HfAlOx/CNT) exhibits both nonvolatile resistive switching (synapse) and volatile threshold switching (neuron) by tuning compliance current. - Synaptic (nonvolatile) mode: Demonstrated resistive switching with compliance currents of 100 µA and 1 mA; endurance and retention up to 1×10^5 s with on/off ratio ~10^3. Stable operation under different straining states and uniform HRS/LRS distributions across 30 devices. - EPSC increases with pulse amplitude (10 ms pulses at 1.5–3.5 V), showing transition from short-term to long-term memory. PPF and PPD observed for paired spikes with 1 s interval and fitted by double exponential decay. - Array-level memory: A 9×9 array stored letters “L”, “T”, and “M” for 1000 s and was erased using −2.5 V, 50 ms pulses. Multi-level conductance: six stable states with retention over 100 s at 100 mV. - Neuron (volatile) mode: With compliance currents of 10 µA, 1 µA, and 100 nA, devices showed volatile threshold switching. Integrate-and-fire behavior demonstrated under pulses of 1 ms: 0.5 V produced no firing; 1.0 V and 1.5 V produced spike responses with strength-modulated frequency; firing threshold defined at 1 pA. - Mechanism: C-AFM current maps at 0, 5, and 10 V indicate gradual growth and dispersion of conductive channels; switching behavior modeled by Ohmic, trap-limited, and trap-filled SCLC regimes. Reconfigurability arises from modulation of Ag filament strength (weak for volatile, strong for nonvolatile). - Energy consumption: Firing energy reduced to 1.9 fJ/spike (E = V × Ifiring × t), at least three orders of magnitude lower than biological neurons (pJ/spike) and below many state-of-the-art artificial neurons (oxide, 2D, others). - System demo: An intelligent heating textile was realized by integrating synapse, neuron, and CNT heater. Synaptic weight modulation adjusted voltage division to neurons, converting spiking responses into heating time/period. Neuromorphic units operated in 60 s cycles; heating unit executed operations with a 3000 s period, demonstrating closed-loop, low-power smart textile functionality.

The work addresses key barriers in neuromorphic textiles by unifying synaptic and neuronal functions within a single fiber-based memristor through reconfigurable switching. This eliminates material and circuit mismatches between separate synapse and neuron devices and simplifies neuron implementation from multi-component circuits to a single element. The demonstrated nonvolatile multi-level synaptic plasticity with high on/off ratio, endurance, and array-level storage supports ANN-style weight storage, while volatile threshold switching enables SNN-like integrate-and-fire computation. The ultralow 1.9 fJ/spike energy consumption far surpasses biological energy efficiency benchmarks and previously reported artificial neurons, highlighting potential for scalable, battery-friendly wearable neuromorphic systems. The integrated heating textile showcases functional coupling of synapse-neuron chains to downstream actuators, validating system-level feasibility of in-textile neuromorphic control for smart garments.

This study introduces a reconfigurable fiber-based memristor (Ag/MoS2/HfAlOx/CNT) capable of operating as both an artificial synapse and artificial neuron within a woven textile architecture. It achieves robust nonvolatile multi-level synaptic storage and volatile integrate-and-fire spiking in a single device, with record-low firing energy of 1.9 fJ/spike. The integration of synapses, neurons, and CNT heaters demonstrates an ultralow-power smart textile capable of autonomous heating control. Future work could focus on scaling to larger textile arrays with on-body testing, enhancing multi-level precision and retention, long-term mechanical durability under wear conditions, closed-loop sensing/actuation for broader functions, and hardware learning demonstrations in realistic tasks.

• Mode control depends on precise compliance current settings, which may complicate large-scale, uniform operation without careful circuitry. - Array demonstrations are limited in size (9×9) and function (image retention); full-scale neuromorphic learning benchmarks and on-chip training accuracy are not reported. - Long-term mechanical durability under repeated bending, stretching, washing, and real wearable conditions is not extensively characterized. - Neuron energy is reported under specific pulse conditions with a very low firing threshold (1 pA); cross-platform comparisons may be sensitive to measurement definitions and conditions. - System demonstration focuses on heating; broader multifunctional closed-loop applications and power budgets under continuous wearable operation remain to be explored.