An ultrasmall organic synapse for neuromorphic computing

The development of electronic gadgets capable of mimicking human-like thinking, decision-making, and interaction is a significant goal in artificial intelligence. Biological neural systems achieve this through complex interactions between biomolecules (neurotransmitters) and metal ions within neurons and synapses. Emulating this molecular-level computing power using advanced semiconductor devices and microelectronic chips is crucial for developing brain-inspired intelligence. Organic and nanomaterial-based synapses, benefiting from their inherent nanoscale properties, versatile electrical characteristics, biocompatibility, and mechanical softness, are key components for neuromorphic computing. To achieve effective neuromorphic computing, research has focused on understanding and controlling the working mechanisms of organic synaptic devices, including charge transfer, redox transitions, structural reconfigurations, and ion migration. A critical challenge is developing novel organic materials that allow for high-density integration of miniaturized devices, mirroring the complexity of biological neural networks. Previous research has explored molecular redox transitions and single-molecule neuromorphic devices, but these demonstrations often lack the necessary scalability and interconnectivity required for real-world chip applications. Efforts to optimize lithography techniques have yielded sub-micrometer organic synapses, but the inherent structural inhomogeneity of most organic materials (due to asymmetric chromophores and weak intermolecular forces) leads to uneven charge carrier processes and hinders nanoscale device fabrication with reliable performance. This study focuses on establishing a structure-property relationship between molecular arrangements, multi-scale ordering, and electronic events to create ultrasmall synapses for high-performance neuromorphic computing. High-performance solid-state materials tend to exhibit ordered structures, and this research introduces a rationally designed organic macromolecule to facilitate the creation of ultrasmall organic neuromorphic devices. The polymer PBFCL10 incorporates oxygen-containing moieties to enhance ion-molecule interactions, fine-tuning electrical behavior through ion migration and conductive filament (CF) evolution. The ordered nanoscale polymeric grains, formed by strong π-π interactions of rigid furan chromophores, regulate the size, spacing, and arrangement of metal nanofilaments. This approach allows for the fabrication of organic synapses with unprecedentedly small dimensions (50 nm) and high integration density (1 Kb).

Extensive research has explored various approaches to develop organic and hybrid resistive switching materials and devices for neuromorphic computing. Studies have investigated different working mechanisms, including charge transfer, redox transitions, structure reconfigurations, and ion migration, to control the electrical behavior of single molecules and their assemblies. Significant advancements have been made in understanding how these mechanisms contribute to the switching characteristics of organic memristors, including the correlation between molecular redox transitions and switching behavior in micrometer-scale devices. Moreover, molecular structure reconfiguration-based single-molecule neuromorphic devices have been demonstrated. However, a major challenge has been achieving scalable integration of these devices with sufficient size and interconnectivity for functional applications. Researchers have also explored methods to optimize lithography techniques to reduce the size of organic synapses to the sub-micrometer range. Nevertheless, the inherent structural inhomogeneity of many organic materials continues to pose a significant obstacle to creating ultrasmall and reliable nanoscale devices. The stochastic nature of ion migration and conductive filament evolution in structurally disordered materials often leads to poor device uniformity and reliability, making it difficult to shrink their dimensions to the nanometer scale and integrate them into high-density arrays for practical applications. Previous work has explored strategies such as dislocation engineering to improve switching reproducibility in inorganic memory devices; however, the translation of these strategies to organic systems requires careful consideration of the unique properties of organic materials and their processing techniques.

The researchers designed a semicrystalline macromolecule, poly(butylene furandicarboxylate)90-b-(ε-caprolactone)10 (PBFCL10), as the switching matrix for the ion-based organic memristive synapse. The polymer incorporates oxygen-containing moieties to facilitate ion migration and conductive filament formation, while rigid furan segments provide molecular crystallinity for ordered thin-film formation. The incorporation of flexible ε-caprolactone components improves solution processability and mechanical flexibility. A 150 nm thin film of PBFCL10 was fabricated on an Ag/SiO2/Si substrate using spin-coating. GIWAXS analysis confirmed the ordered molecular stacking and in-plane alignment of lamellar crystals, which facilitated the formation of an ordered network for metal ion migration along the polymer grain boundaries. AFM observations revealed smooth surface morphology with orderly aligned fibrillar grains. C-AFM measurements showed a dense distribution of high-current regions, indicative of conductive filaments, with an estimated inter-spacing of 40-50 nm. The researchers fabricated a 32 x 32 crossbar array of ultrasmall organic synaptic devices with a linewidth of 50 nm and separation of 85 nm. Electrical characterization revealed linear switching between 32 conductance states with high uniformity (cycle-to-cycle: 98.89%, device-to-device: 99.71%). A mixed-signal neuromorphic hardware system, incorporating the PBFCL10 synapse array and an FPGA controller, was implemented to execute spiking-plasticity-enabled algorithms for decision-making tasks. A Hopfield neural network (HNN) was used for travel planning tasks, and the convergence of the network was optimized using a biexponential function derived from the spike-rate-dependent plasticity (SRDP) dynamics of the organic device. The performance of the system was assessed by comparing the iteration efficiency and accuracy of different HNN variants. The fabrication of the nanoscale Au/PBFCL10/Ag devices in 32 x 32 arrays involved electron beam lithography and a lift-off approach. Electrical measurements were performed using a Keithley 4200 semiconductor parameter analyzer. The integration of the synapse crossbar array into the hardware system involved wire bonding, encapsulation, and connection to a PCB-FPGA control system, which included a DAC, an ADC, a MUX, TIAs, a power voltage source, and an I/O header. The FPGA (Altera Cyclone IV) controlled the voltage pulses applied to the devices and monitored their conductance states. A PC-based user interface allowed for parameter setting and monitoring of the device operation.

This research successfully demonstrated an ultrasmall organic synapse based on the semicrystalline polymer PBFCL10. The key findings include: 1. **Ultrasmall Device Dimensions:** The fabricated synapses achieved the smallest device dimension reported to date, measuring only 50 nm. This is a significant reduction compared to previous organic synapses. 2. **High Integration Density:** The researchers created a neuromatrix with a high integration size of 1 Kb (32 x 32 array), demonstrating the scalability of the proposed approach. 3. **Linear Conductance Switching:** The PBFCL10 synapses exhibited linear switching between 32 conductance states, offering fine-grained control over synaptic weights. 4. **High Uniformity:** The devices showed exceptionally high uniformity, both cycle-to-cycle (98.89%) and device-to-device (99.71%), indicating highly reliable performance. This is a significant improvement over previous organic synaptic devices. 5. **Functional Neuromorphic System:** A mixed-signal neuromorphic hardware system was built, integrating the organic neuromatrix with an FPGA controller. This system successfully executed a spiking-plasticity-related algorithm for decision-making tasks. 6. **Effective Travel Planning:** The system implemented a modified Hopfield neural network to perform travel planning tasks and consistently found the optimal routes. 7. **Superior Annealing Algorithm:** The study also highlighted the superiority of the spike-rate-dependent plasticity (SRDP) annealing algorithm compared to linear and exponential annealing approaches. The SRDP algorithm improved both the accuracy (by 16.7%) and efficiency (by 31.7%) of the route finding task. The SRDP's dynamic modulation of the damping rate during iteration surpasses the fixed rates of linear and exponential annealing. 8. **Structural Stability:** GIWAXS measurements confirmed the stability of the PBFCL10 thin film's crystallinity and orientation even after repeated conductive filament formation and rupture cycles, indicating excellent device durability and longevity. This stability is crucial for practical applications.

The results of this study address the critical challenges of miniaturization, scalability, and reliability in organic neuromorphic devices. The use of the rationally designed PBFCL10 polymer, with its ordered structure and ability to facilitate uniform conductive filament formation, enabled the fabrication of ultrasmall, highly uniform, and reliable organic synapses. The successful implementation of a mixed-signal neuromorphic hardware system demonstrates the potential for using these devices in practical applications. The high uniformity and linear conductance switching characteristics of the PBFCL10 synapses are crucial for accurate and reliable neuromorphic computing. The superior performance of the SRDP-based annealing algorithm compared to linear and exponential annealing demonstrates the potential for using bio-inspired algorithms to improve the efficiency and accuracy of neuromorphic computing tasks. The high integration density achieved (1 Kb) is a major step towards creating larger-scale neuromorphic systems capable of performing complex computations. The findings have significant implications for the development of energy-efficient, bio-inspired computing systems.

This research successfully demonstrated an ultrasmall organic synapse with superior performance characteristics, paving the way for high-density neuromorphic computing. The use of the rationally designed PBFCL10 polymer yielded devices with unprecedentedly small dimensions, high uniformity, and linear conductance switching. The integration of these devices into a mixed-signal neuromorphic system demonstrated their functionality for decision-making tasks. Future research could explore different polymer architectures, optimization of device fabrication methods, and integration of these synapses with other neuromorphic components to create more complex and powerful neuromorphic systems. Investigating the potential of using native ions within the organic material to form filaments could lead to further improvements in energy efficiency and scalability. The development of flexible and stretchable organic neuromorphic systems based on similar principles could open up new applications in bioelectronics and human-machine interfaces.

While the study demonstrates significant advancements in organic neuromorphic devices, some limitations exist. The current design utilizes a blanket organic insulator, and further patterning of the switching medium using lithographic techniques could improve device performance and reduce cross-talk. The study focused on a specific type of neural network (Hopfield network) for the travel planning task. The generalizability of the findings to other types of neural networks and tasks should be investigated. The current system relies on external metal ions for conductive filament formation, and the use of intrinsic ions from the polymer itself may enhance the long-term stability and power efficiency of the devices.