An ultrasmall organic synapse for neuromorphic computing

Neuromorphic electronics aim to emulate the molecular-level computing power of biological neural systems, where information processing relies on interactions among neurotransmitters and ions within neurons and synapses. Organic and nanomaterial-based synapses are attractive due to their biomimetic electrical behavior, biocompatibility, and mechanical softness. However, in organic devices, structural inhomogeneity and lack of long-range order often yield spatially uneven charge generation, transport, and collection, leading to poor uniformity and limited scalability. To enable dense, nanoscale integration with reliable performance, it is necessary to understand and control structure–property relationships—particularly molecular arrangements, multiscale ordering, and ion migration pathways—to realize ultrasmall, high-performance organic synapses.

Prior efforts to realize organic neuromorphic devices have explored mechanisms including charge transfer, redox transitions, structural reconfiguration, and ion migration to modulate electronic states. In-operando spectroscopy and scanning probe studies have correlated molecular redox transitions with switching behavior in micrometer-scale organic memristors, suggesting feasibility for nanoscale devices. Single-molecule neuromorphic devices based on structural reconfiguration have also been reported. Parallel to materials advances, lithography optimizations have reduced electrode linewidths of organic synapses to sub-micrometer scales. Nevertheless, due to inherent material disorder, stochastic ion migration and filament evolution have limited uniformity and scalability. Dislocation engineering in inorganic systems (e.g., SiGe epitaxial memory) has inspired strategies to improve switching reproducibility. Building on this context, the present work develops a crystalline-control approach using a biomass-derived, structurally ordered polymer to homogenize metal nanofilament formation in organic switching media for scalable, reliable nanoscale integration.

Materials: A semicrystalline block copolymer, poly(butylene furandicarboxylate)90-b-(ε-caprolactone)10 (PBFCL10), was synthesized via polycondensation from poly(ε-caprolactone) diol oligomers and bio-based 2,5-furandicarboxylic acid. The design incorporates oxygen-containing moieties for metal cation migration and rigid furan segments to induce crystallinity, with flexible ε-caprolactone segments for processability and mechanical compliance. Thin-film formation and structural characterization: A ~150 nm PBFCL10 film was spin-coated onto Ag/SiO2/Si substrates. GIWAXS revealed ordered molecular stacking with peaks at 2θ ≈ 18.3° and 21.3°, attributed to (001) and (010) planes of furan-based lamellar crystals. Spin-coating induced in-plane vortex alignment of lamellae; thermal annealing at 60 °C promoted out-of-plane ordering. AFM showed smooth morphology (rms ~0.37 nm) with aligned fibrillar grains (~45 nm diameter). Conductive AFM on ON-state regions revealed dense, quasi-bulk conductive hotspots consistent with conductive filament (CF) formation along grain boundaries, with inter-filament spacing ~40–50 nm and local high-current regions up to ~40 nm in small-area scans. GIWAXS before/after SET/RESET indicated stable crystallinity and orientation during CF formation/rupture cycles. Device fabrication and electrical testing: Nanoscale Au/PBFCL10/Ag memristive synapses were patterned into 32 × 32 crossbar arrays via electron-beam lithography and lift-off. The minimum electrode linewidth was 50 nm with 85 nm separation. Electrical measurements used a Keithley 4200. Devices exhibited single-Ag CF formation within 50 nm cells, enabling controlled filament evolution. Negative-feedback RESET produced atomic point contact (APC) structures, yielding stepwise conductance quantization and linear multilevel modulation. Hardware system and algorithm: The crossbar array was wire-bonded and packaged into 64-pin chips, encapsulated with a 1.9 cm × 1.9 cm quartz plate, and integrated on a PCB with an FPGA controller (Altera Cyclone IV), DAC, ADC, MUX, TIAs, power supply, and I/O. A PC-based interface controlled programming (voltages, compliance currents, pulse counts) and monitored conductance evolution. For in-memory computing (VMM operations), a lightweight Hopfield neural network (HNN) framework was implemented entirely on the mixed-signal platform. A spike-rate-dependent plasticity (SRDP)-based biexponential annealing (BHNN) was used to dynamically adjust synaptic weight updates and enhance convergence for decision-making/travel-planning tasks. Conductance matrices were updated in-operando and compared to theoretical targets throughout iterations.

- Material/system scaling: Achieved ultrasmall organic synapses with 50 nm device size and 85 nm line separation in a high-density 32 × 32 (1 Kb) crossbar array. - Conductance modulation: Devices exhibited 32 quantized conductance states with linear evolution; quantized conductance showed a high linear fit (R^2 = 0.997). APC formation via negative-feedback RESET enabled stable, stepwise modulation. - Uniformity and variability: Demonstrated high uniformity—98.89% cycle-to-cycle and 99.71% device-to-device. Minimum conductance variation reported down to 0.29% for quantized states. - Structural robustness: GIWAXS before/after switching showed negligible change, indicating stable crystallinity and orientation during CF cycling. - In-operando neuromorphic computing: Mixed-signal hardware executed iterative HNN tasks; the in-operando updated 32 × 32 conductance matrix matched the theoretical BHNN-updated matrix with 99.77% proximity. - Decision-making performance: BHNN found a shortest-path route with travel time ≈4.166. Compared to linear and exponential annealing, BHNN achieved higher accuracy (96.7% vs 60% and 80%) and faster convergence (<420 vs 800 and 615 epochs), corresponding to ~31.7% efficiency gain over exponential annealing and 16.7% accuracy improvement over it. - Filament engineering: Single-Ag CF per 50 nm cell and controlled APC formation underpin reliable bistable switching and precise multilevel weight updates.

By engineering a semicrystalline polymer matrix with ordered molecular packing and aligned lamellar grains, the authors homogenize ion migration pathways and CF nucleation along grain boundaries. This suppresses the stochasticity common in disordered organic media, enabling reliable nanoscale devices and dense crossbar integration. The controlled formation and refinement of single Ag filaments—and APCs under negative-feedback RESET—provide a physical basis for linear, quantized multilevel conductance, directly addressing the need for precise, repeatable synaptic weight updates in neuromorphic computing. The mixed-signal system demonstrates that organic synapse arrays can be integrated with standard digital control (FPGA) to perform in-memory vector–matrix operations and iterative optimization. The SRDP-based BHNN annealing leverages device dynamics to enhance convergence speed and solution quality, validating the practical utility of the material–device–algorithm co-design. Overall, the findings show that structural order at the molecular and mesoscopic scales is pivotal for scaling organic neuromorphic devices while maintaining uniformity, reliability, and algorithmic performance.

This work introduces a crystalline-control strategy in a biomass-derived semicrystalline polymer (PBFCL10) to realize ultrasmall, uniform organic synapses. The approach yields 50 nm devices in a 32 × 32 crossbar array with linear 32-state conductance quantization, excellent uniformity, and robust structural stability. A mixed-signal neuromorphic hardware platform integrating the organic neuromatrix and FPGA executes SRDP-informed BHNN algorithms for decision-making tasks with high accuracy and efficiency. Future directions include: (i) lithographic patterning/etching of the organic switching medium to isolate devices for further performance gains; (ii) replacing electrode-injected mobile metal ions with native ions in the polymer to form semiconducting filaments, potentially reducing current and energy per operation; and (iii) developing fully flexible, all-organic neuromorphic systems on soft/stretchable substrates to enable bioelectronic interfaces and wearable/implantable applications.

The present crossbar employs a blanket organic insulator rather than individually patterned device islands, which may limit isolation and optimal performance. Filament formation relies on electrode-injected Ag+ ions; transitioning to native ionic species could reduce power but requires balancing low-power operation with non-volatility for reliable weight retention. The demonstrated system is an organic–silicon hybrid relying on digital control and algorithmic annealing; fully organic implementations and comprehensive energy/retention benchmarking remain for future work.