Artificial synapse based on carbon quantum dots dispersed in indigo molecular layer for neuromorphic applications

The technological advancement often associated with the development of computing systems is severely affected by its architecture when the data processing becomes time-consuming, especially in the case of “big data” analysis. The limitation is that the information processing unit and the memory reside at two physically distinct locations in the hardware of computing systems. Meanwhile, there is an exponential growth in data availability, which demands the capability of storing and processing gigantic amounts of complex information in an accelerated way. A new and innovative design of electronic circuitry is needed to satisfy the demands of the enormous amount of data available to process and save complex information. This will also facilitate deep learning and artificial intelligence (AI) with the use of low-power-consuming device architectures.^1–4 The working of a normal human brain can be taken as a model system in such cases of massive information processing with a very low power consumption (∼20 W) and a very good efficiency and longevity.^5–8 Often, the human brain outperforms the state-of-the-art supercomputers in a big way, with an added advantage of decision-making. The brain-inspired device architecture attracts scientists as it comprises massive parallel information processing with better energy efficiency compared to modern computers.^9,10 Recently, many activities, such as speech recognition, image processing, and decision-making based on artificial intelligence, have taken giant leaps in technological advancement.^11,12 This kind of AI-based technology, indeed, requires superior computing power, where the traditional complementary metal–oxide–semiconductor (CMOS)-based architecture renders no help as CMOS-based computing systems are hitting the roadblock because of the difficulty in scaling of devices up to brain-like complexity.^3,13,14 The main objective of brain-inspired or neuromorphic devices and computation is to develop spiking neural networks (SNNs) using artificial neurons and synapses. Voltage pulses are applied across the synapse, and post-synaptic current (PSC) will be monitored across the synapse, mimicking the effect of an action potential in the biological synapses.^15,16 The artificial synapse will give a weighted output similar to the biological synaptic output.

Brain-inspired computing devices mainly consist of pre-synaptic neuron/synapse/post-synaptic neuron structures, which normally act like two-terminal metal/switching material(s)/metal devices.^17–20 The two-terminal devices mimic the fundamental activities of devices, such as learning and forgetting, normally referred to as potentiation and depression.² The core of such devices is the switching materials that are carefully sandwiched between the two metallic electrodes. Mainly, memristors have consistently proved to be ideal devices that can perform/show fundamental learning and forgetting characteristics.^21–23 Many oxides (thin films and nanoparticles) and their derivatives have been tested for synaptic applications using a memristive action with both two-terminal and three-terminal device structures.^24,25 Due to the possibility of a flexible device architecture, organic materials as resistive switching materials have also attracted a lot of attention for fabricating artificial synapses.^26–30 In particular, the highly conductive polymer poly(3,4-ethylenedioxythiophene):poly(styrene sulfonate) (PEDOT:PSS) with tunable flexibility and conductivity has attracted much attention.^31–33 Most of the studies on artificial synapses involve applying voltage pulses or spikes. Recently, there have been reports on optical spike-based artificial synapses as well. The post-synaptic current (PSC) will be monitored after applying appropriate optical pulses. The optical pulse-based technique is a non-invasive technique with long-distance operation capabilities.⁴ Furthermore, the combination of electrical and optical pulses could also be used for the multi-functionality of the artificial synapses. The molecular-based devices have tremendous advantages in fabricating artificial synapses due to their cost-effective and bio-compatible properties. It has been proven many times that polymer-based devices with inclusions such as nanodots,^2,34 nanotubes,³⁵ and nanosheets³⁶ have proved to be much more efficient. The inclusions act as charge-trapping centers that facilitate resistive switching (RS). On the other hand, analog-type resistive switching memory devices will also find importance due to the possibility of continuous variation of resistance states to realize synaptic functions. We use zero-dimensional carbon quantum dots as inclusions to the indigo molecular matrix. The carbon quantum dots (CQDs) are prepared using edible jaggery. It is known that the surface defects in carbon quantum dots can control the generation and recombination of electron–hole pairs, thus controlling the functioning of the artificial synapse. Herein, we report a detailed analysis of an artificial synaptic device based on carbon quantum dots dispersed in the indigo molecular layer stable at room temperature. The surface defects on the CQDs act like trapping centers that will promote the conduction path in the devices, thus, stabilizing the different resistance states. A detailed study on learning and forgetting functionality (A200–indigo) is achieved at room temperature.

In this study, we demonstrate the synaptic characteristics of a two-terminal device consisting of carbon-quantum dots embedded in the indigo molecular layer. We prepared the carbon-quantum dots (CQDs) from commercially available jaggery. CQDs were synthesized by the thermal decomposition method.^37–44 A certain amount of jaggery was poured into a crucible and annealed in a box furnace at 200, 220, and 240 °C for four hours each. The crucible was allowed to cool to ambient temperature after annealing. The product we received after annealing had a dark brown color, and we crushed it into powder using a mortar and pestle. 1.05 g of the obtained powder was added to 30 ml of methanol, and the solution was ultrasonicated for two hours. The dark brown suspension was filtered out by using two Whatman filter papers (diameter 125 mm and pore size 11 µm). For further characterization and device fabrication, the filtered CQDs (dispersed in methanol) were transferred to a glass vial and preserved. Since the precursors were annealed at three different temperatures, 200, 220, and 240 °C, the CQDs are named A200, A220, and A240 and will be denoted as such throughout the rest of the article. The synthesis procedure of CQDs has been optimized to consistently produce CQDs around 3.5–4.5 nm in size, as revealed by transmission microscopy imaging shown in Fig. 2(a). A large-scale TEM image and histogram are plotted in Fig. S1 (supplementary material). CQDs prepared at 200 °C showed lattice fringes with a typical d-spacing of 0.196 nm, which we attribute to (100) graphitic planes, while the CQDs prepared at 240 °C show a lattice spacing of 0.288 nm, typical of (020) graphitic planes. Interestingly, the literature says that these are single crystalline carbon quantum dots.^45–47 We measured the optical energy gap of these CQDs via UV–vis spectroscopy, as shown in Fig. 1(c). The CQDs show a systematic change in the energy gap depending on the annealing temperature of the precursor material. We believe that the higher annealing temperature has reduced the size range of the carbon quantum dots. A similar observation was also reported that the average size has reduced due to the annealing temperature.⁴⁸  The room temperature photoluminescence reveals a broad emission spectrum with maxima around 500 nm, showing the green color emission from the quantum dots. A red shift in the emission spectrum as a function of the excitation wavelength is shown in Fig. 2(e). This is purely a quantum confinement effect due to the size distribution of the CQDs. The red shift of the emission peaks from the CQDs has in itself been extensively studied.^49–51 All the CQDs prepared show a similar trend of red shift in the emission spectra with the excitation wavelength, as shown in Fig. 2(f). We have checked the photoluminescence properties of A220 and A240 CQDs as shown in Fig. S2 (supplementary material). For the device fabrication and all the discussion in the following, we use A200 with indigo molecules. For the indigo molecules, Fourier-Transform infrared spectroscopy (FTIR) spectroscopy reveals pronounced signatures of the side groups, such as the N–H bond stretching at 3269 cm⁻¹, the C=O and C–C stretching modes at 1625 cm⁻¹, and the C–H bending modes at 750 cm⁻¹. The FT-IR spectrum is shown in Fig. 2(d). These values match very well with the modes of the indigo molecule available in the literature.^52,53

An indigo molecular solution was made by combining 0.3 g of indigo molecules with 7 ml of methanol, followed by ultrasonication for 30 min. Furthermore, the already prepared CQDs/methanol solution is mixed, and finally, solutions are ultrasonicated for 30 min to get a uniform distribution of the CQDs in the indigo molecules. We use pre-fabricated Pt electrodes (MICRUX Technologies) with a 10 μm separation. The device configuration is shown in Fig. 2(b). The prepared A200–indigo solution was drop-cast and dried at 100 °C for 4–5 h. The electrical conductivity measurements are done in both dc-voltage sweep and pulsed mode measurements using a Keysight B2902B source meter. These two-probe measurements are done at room temperature under ambient conditions.

With a detailed understanding of the prepared CQDs and the indigo molecules, we fabricated two-terminal Pt/(A200–indigo)/Pt devices.⁵⁴ A typical two-probe current–voltage characteristic is shown in Fig. 3(a). The scanning direction is indicated by the arrows. The voltage scanning sequence is (−5 V)–(0 V)–(+5 V)–(0 V)–(−5 V). The dotted lines indicate the high resistance state (HRS), and the continuous line indicates the low resistance state (LRS). All the current–voltage scans show the continuous transition from the LRS to the HRS state associated with negative differential resistance (NDR) characteristics.^25,55–57 The peak value of the NDR occurs at −1.55 V and +1.66 V, respectively, in opposite scanning directions. The current change measured as the peak-to-valley ratio ranges between 4% and 7%.

Before going to the synaptic characteristics of our devices, let us examine the conduction mechanism in the A200–indigo molecular device in detail. To check the occurrence of the current–voltage curves and associated NDR features, we repeat the current–voltage measurements as shown in Fig. 3(a). The device shows extremely stable and consistent current–voltage characteristics. The negative differential resistance region showing up in each scan indicates that it is a characteristic feature of such devices.⁵⁸ We take the average curve for analysis as shown in Fig. 3(b). Since our device is a symmetric device, the symmetric occurrence of the NDR feature is understandable. It is interesting to note that the NDR occurs only in samples with inclusions of CQDs, as shown in Fig. S3 (in the supplementary material). For detailed analysis, we consider the positive bias region in detail. As shown in Fig. 3(b), the region can be divided into five regions, including the NDR region. Different regions can be modeled to different conduction mechanisms. Region I is characterized by electron conduction via thermionic emission, as shown by the linear log I vs V^0.5 plot. The charge carriers are from the defects in the A200–indigo switching layer. These defects can come from the bulk defects in the indigo molecular layer. As the bias voltage increases beyond the plateau of the NDR, more and more carriers are injected from the electrode. Region III shows a linear log I–log V plot, indicating ohmic conduction, and we designate this region as the ON state. While scanning the voltage back toward zero, the regions can be divided into two regions, as indicated in Fig. 3(b). Due to the intrinsic repulsion between the charge carriers, a space charge is built near the Pt/A200–indigo interface. This effect further reduces the injection of charge carriers into the switching layer. This effect changes the conduction mechanism from ohmic conduction to space charge-limited conduction (SCLC). In this region, the current is directly proportional to the square of the voltage, and this region is indicated as region IV in Fig. 3(e). The current due to the injected carriers is drastically reduced due to the presence of traps in the switching layer. Eventually, the injected charge carriers fill out all the traps, and then, the trap-charge limited conduction dominates, which has an exponential relation to the applied bias voltage. This is also assisted by the field emission at higher bias voltages.^59,60 As indicated in Fig. 3(f), for lower voltages, the conduction is pure via tunneling as there is not enough energy available for the injected electrons to overcome the barrier. For very low voltages, the barrier is approximated as a rectangular barrier. As the applied bias is increased, the shape of the barrier changes slowly, and for large enough applied voltages, it changes into a triangular barrier, as shown schematically in Fig. 3(f) . The electron conduction at low voltages can occur only via direct tunneling, and for very high bias voltages, it occurs via field emission or Fowler–Nordheim tunneling. We see a transition from direct tunneling to field-assisted conduction at around 1.82 V, which is shown as V_trans.^61,62 From this analysis, it is clear that the conduction in A200–indigo occurs via interface-assisted processes in the majority of the bias voltages applied. The variety of processes explained here will result in the hysteric current–voltage characteristics of the devices. The hysteric current–voltage characteristics lead us to the memory application of the Pt/CQD+indigo/Pt devices. Devices with A200, A220, and A240 CQDs are tested for read–write–erase cycles. We use +5 V to write the information and read it at +1 V. Again, we erase the information at −5 V and read it at +1 V. We used 200 ms pulses with respective magnitudes of the voltage. Figures 4(a), 4(c), and 4(e) show the pulse voltages applied during the read–write–erase cycles. The current follows the pulse voltages with differences in the resistance state before and after the write pulse. The difference in the resistance states for 1500 cycles is plotted in Figs. 4(b), 4(d), and 4(f). All devices show an excellent stability, and the HRS and LRS states are well separated with large resistance differences. As indicated, the difference between the HRS and LRS is shown with the double-headed arrows, and it is clear that our devices show excellent HRS and LRS states, indicating the suitability of the CQD-embedded molecular devices for memory applications.^63–68 

In conclusion, we have reported an extremely stable neuromorphic device showing long-term potentiation (LTP) and long-term depression (LTD). The devices are based on carbon quantum dots synthesized via a simple route using an edible source. The ease of synthesis of the room temperature stable carbon quantum dots (CQDs) using edible sources is also an advantage in such applications. The uniform distribution of carbon quantum dots stabilizes the resistive switching devices. The two-terminal devices showed not only memory functions but also excellent learning and forgetting characteristics, which are essential for realizing the artificial synapse for neuromorphic computation working at room temperature. The read–write–erase cycles showed excellent high resistance state (HRS) and low resistance state (LRS) stability for more than 1250 cycles. The usage of a molecular-based system for such applications attracts more attention due to the ease of fabrication of devices and room-temperature operation.

See the supplementary material for more information on the TEM images, size distribution, photoluminescence studies of A220 and A240 CQDs, and multi-cycle potentiation and depression curves.

The authors would like to thank Centre for Functional Materials (CFM), Vellore Institute of Technology, for its support during this research.

The authors have no conflicts to disclose.

Amrita Bharati Mishra: Data curation (equal); Formal analysis (equal); Methodology (equal); Writing – original draft (equal); Writing – review & editing (equal). Ramesh Mohan Thamankar: Conceptualization (lead); Formal analysis (lead); Investigation (lead); Project administration (lead); Supervision (lead); Validation (lead); Writing – original draft (lead); Writing – review & editing (lead).

The data that support the findings of this study are available from the corresponding author upon reasonable request.