Mos2 Synapses Achieve 85% Memory Transition for Arithmetic and Image Recognition

Scientists are increasingly focused on replicating the energy efficiency of the human brain, leading to intense research into artificial synapses that combine data storage and processing. Umakanta Patra, Subhrajit Sikdar, and Roshan Padhan, from the Indian Institute of Technology, Bombay and Jackson State University, alongside Amandeep Kaur, Satyaprakash Sahoo, and Subhabrata Dhar et al., now demonstrate that field-effect transistors built from chemically-grown bilayer molybdenum disulphide (MoS2) effectively mimic biological synaptic function. This research is significant because these devices exhibit key synaptic behaviours , including pair-pulse facilitation and short-to-long term memory transition , with remarkably low energy consumption (as little as 20 nanojoules per event), and crucially, can perform arithmetic operations and achieve image recognition accuracy of around 85% , paving the way for transparent, flexible, and ultra-low power electronics.

MoS2 transistors emulate synaptic function and learning

This breakthrough reveals a pathway towards brain-inspired neuromorphic computing, integrating data storage and processing for significantly reduced power consumption. By fabricating three-terminal field effect transistors, the team overcame limitations found in two-terminal memristive devices, achieving greater stability, repeatability, and control over transport properties. The unique layered structure of bilayer MoS2, with its potential for interfacial intercalation, allows for the creation of tunable trap states, enhancing persistent photoconductivity and enabling precise control over synaptic weight programming. This work opens exciting possibilities for future advancements in artificial intelligence and machine learning, paving the way for more efficient and adaptable computing systems. The ability to manipulate the memory state through gate bias, coupled with the exceptionally low energy consumption, suggests a promising future for 2L-MoS2 in a wide range of applications, from wearable electronics to advanced robotics. The findings underscore the potential of 2D materials to revolutionise the field of neuromorphic computing and address the growing demand for sustainable, high-performance technologies.

MoS2 Transistors Mimic Synaptic Learning Behaviour, offering potential

To fabricate these devices, the team raised a furnace to 700°C at a rate of 10°C/min, maintaining this temperature for 10 minutes while keeping a sulfur-containing boat at approximately 150°C, before allowing natural cooling to room temperature. Following growth, the 2L-MoS2 layers were transferred onto SiO2 coated Si substrates using a polystyrene-based transfer method, a technique detailed in a prior publication. Characterization involved recording optical and atomic force microscopy (AFM) images at room temperature, alongside photoluminescence (PL) and Raman spectroscopy performed using a Renishaw Invia Reflex micro-PL/Raman setup equipped with a 50× magnification objective lens and a 532nm laser at 500 μW power. A 500cm focal-length monochromator, coupled with an 1800 gr/mm grating and CCD detector, captured the spectra.

High resolution transmission electron microscopy (HRTEM) images were obtained using a Thermo Scientific Themis 300 G3 system operating at 300kV, after transferring the film onto a Cu-grid via polystyrene-based wet transfer. The team then fabricated 2L-MoS2 based back-gated FETs using standard optical lithography, followed by sputtering deposition of 30nm of Ti and 100nm of Au. Electrical characteristics were measured with a Keithley 6487 Pico ammeter-voltage source and a Keysight 2B2900 source measuring unit, while a custom-built setup employing a Xenon lamp, monochromator, and associated apparatuses illuminated the devices with varying wavelengths and powers. Raman spectroscopy revealed peaks at 21cm-1 separation, consistent with bilayer MoS2, and PL spectroscopy identified exciton peaks at 1.84 eV and 2 eV. AFM imaging confirmed a film thickness of 1.7nm, while HRTEM imaging and SAED patterns validated the highly crystalline structure of the grown MoS2 films.

MoS2 transistors emulate synaptic function and learning

Experiments revealed that the rate of memory state depression can be precisely controlled by adjusting the gate bias, offering a novel mechanism for memory manipulation. Raman spectroscopy confirmed the quality of the 2L-MoS2 film, with the E2g1 and A1g phonon modes separated by 21cm-1, consistent with reported values for bilayer MoS2. Photoluminescence measurements identified peaks at 1.84 eV and 2 eV, corresponding to neutral A- and B-excitons, further validating the material’s characteristics. Atomic force microscopy (AFM) measurements established a film thickness of 1.7nm, aligning with expectations for bilayer MoS2, while high-resolution transmission electron microscopy (HRTEM) confirmed the highly crystalline structure of the grown films.

These detailed material characterizations provide a strong foundation for the observed synaptic behaviour. The team measured a hysteresis window of approximately 8V in the dual sweep Ids−Vgs profile of the 2L-MoS2 based back-gated FET device. Threshold voltage (Vth) was determined to be −10.2V, calculated from the intercept of the √Ids versus Vgs plot. Electron concentration (n) was calculated using the equation n= Cox× (Vgs−Vth) / e, with the gate oxide capacitance (Cox) estimated to be 11.5 nF/cm2. Background electron concentration at Vgs= 0V was found to be 7.3 × 1011cm-2, and the electron mobility (μ) was calculated as 0.03 cm2V-1s-1. The researchers observed that the devices exhibit resistive switching (RS) behaviour, with changes in slope observed in log-log plots, potentially attributed to sulphur vacancy migration, mirroring neurotransmitter activity in biological systems. This work demonstrates a great prospect for 2L-MoS2 in developing low power, transparent and flexible devices.

MoS2 transistors emulate synaptic plasticity and efficiency remarkably

Furthermore, the study quantified exceptionally low energy consumption, 280 fJ for electrical and 20 nJ for optical stimulation per synaptic event, highlighting the potential for energy-efficient computing. The authors acknowledge a limitation in the scope of their ANN simulation, but suggest future work could explore more complex network architectures and larger datasets. These findings establish a promising pathway towards developing low-power, transparent, and flexible devices for neuromorphic computing applications.