Emulating Ebbinghaus forgetting behavior in a neuromorphic device based on low dimensional h-BN

The design and fabrication of optoelectronic devices performing in the ultraviolet (UV) range are crucial for military applications.¹ Wide-band semiconductors are considered the best candidates for such applications. Many materials such as AlN and AlGaN have been tested for the UV-photodetection process.^2,3 Recently, photo-detectors using 2D materials with a wide bandgap have also been reported. Mainly, monolayer hexagonal boron nitride (h-BN) is very interesting because of its large energy gap (∼6 eV) and graphene-like structure.^4–6 The heterostructure containing graphene/h-BN/ZnO has also been used as a UV-photodetector.⁷ Resistive switching based on h-BN is established and shows extremely stable performances. Threshold switching is shown by Au/h-BN/Au devices, offering very low switching voltage (<1 V) and ultra-low energy consumption down to zeptojoule.⁸ To understand the electrical breakdown when used as a dielectric, conduction atomic force microscopy has been utilized to understand the defect generation and its impact on the dielectric breakdown.⁹ h-BN based synaptic devices are an exciting platform to address the challenges in miniaturization and energy consumption in next-generation computing. They can mimic the human visual and memory systems, which can process visual signals and store the data locally.^10–14 Persistent photoconductivity (PPC) in optoelectronic devices is analogous to the generation of post-synaptic current (PSC) in post-synaptic neurons when neurotransmitters are generated from pre-synaptic neurons in the biological synapse. The photocurrent in optoelectronic synapses can be controlled by exposure time, intensity, frequency, and the number of pulses, and it is critical to understand the influence of each of them in a synaptic behavior to optimize the performance of devices.

Synaptic devices that are sensitive to UV photons and have processing and storage capability can improve the device complexity.¹⁵ A fast response is desirable for increasing the operating speed. Optical signaling helps in achieving a higher bandwidth and decoupling from electronic noise.¹⁶ In optical signaling, light acts as an additional degree of freedom for the device to adjust the synaptic properties. It will be an important development in hardware design to construct human brain-like computing systems that can emulate the learning and forgetting processes happening in the human brain. More importantly, the human brain has well-developed sensory organs that can sense external stimuli such as light, pressure, and chemical molecules. The artificial neuromorphic systems should have such characteristics. Furthermore, it is established that almost 80 percent of the information perceived by the human brain is via the visual system.^13,17 The perceived information is stored in memory and often lost in the form of a forgetting process. The memory and learning process is controlled by the synapse in the biological neural network. Various materials have been studied for photonic synapses, such as oxide semiconductors, binary oxides, low dimensional materials in the form of 0D, 1D, and 2D, organic materials, and heterostructures.^18–22 They require less energy per spike, enabling faster response than electrical counterparts. There are significant advantages for devices that can be switched by optical means since optical signaling supports higher bandwidth communication and faster transmission speeds. Some of the potential applications include switchable memory elements in optical communications and photonic integrated circuits, optically tunable synaptic elements, and new types of light sensors that can be used in cellular neural network cameras.^23–25 In-memory processing of the image can be performed directly at the level of the focal plane on the camera chip, which avoids the slow data transfer bottleneck between the optical sensor and the main processor.²⁶ 

Due to the presence of intrinsic atomic defects in monolayer h-BN, often, the theoretical bandgap of $∼6$ eV is not achieved. Defects in h-BN can act as a single photon source, and many defects have been proposed to emit photons at various wavelength ranges.^27,28 Defects such as nitrogen-vacancy (V_N), boron vacancy(V_B), and anti-site defect complex (nitrogen occupying the boron site near a nitrogen-vacancy N_BV_N) are common in h-BN.²⁷ Boron vacancy is theoretically predicted to emit photons in the ultraviolet range, while nitrogen vacancies and the anti-site defect complex are found to emit photons in optical wavelengths.²⁹ 

Among all the devices, memristive devices have been showing great potential to emulate the biological synaptic functionalities. Learning-forgetting-relearning synaptic functions have been studied well in memristive structures involving different materials, and heterostructures. Low-dimensional materials such as graphene, MoS₂, and MoO₃ have been used to emulate the synaptic functionalities.^33–35 Hexagonal boron nitride is a commonly used two-dimensional material for resistive switching devices, sensors, and memristors and is sensitive to ultraviolet light and exhibits strong photoluminescence in the UV-region.^8,36–40 We have fabricated a lateral optical synaptic device using low dimensional h-BN to study the conductance decay properties. We used the light of 375 nm UV light for optical stimulation, focusing on the effect of exposure time, frequency of pulses, and number of pulses. Synaptic behavior is achieved using persistent photoconductivity (PPC), which is a light-induced photoconductivity where the photogenerated charge carriers persist after the light excitation is terminated. The observed decay can be attributed to the defect states (defects in the h-BN lattice), which trap the charge carriers and delay the recombination of photogenerated carriers.^41–43 

A planar Pt/h-BN/Pt device is fabricated using synthesized h-BN. 20 mg of synthesized h-BN is ultrasonicated in 10 ml methanol for 30 min to exfoliate the layers. Complete details of the synthesis of monolayer h-BN are explained in detail in our previous publication.⁴³ The UV–vis spectra shown in Fig. 1(a) indicate high absorbance near 375 nm, prompting us to use ultraviolet light (∼375 nm) for synaptic measurements. To confirm the presence of single-layer h-BN in our device, transmission electron microscopy of h-BN was performed. Figure 1(b) shows the atomically resolved TEM image, and the corresponding diffraction image shown in Fig. 1(c) confirms the presence of single layer h-BN. Interdigitated platinum electrodes from Micrux Technologies are cleaned by ultrasonication in methanol and DI water for 10 min, respectively. The schematic diagram is shown in Fig. 1(d). 10 μl of the sonicated solution is drop-cast into a platinum interdigitated electrode (IDE6-Pt) having a gap of 5 μm and electrode width of 10 μm. This is then dried at 120 °C overnight. Measurements are carried out using a Keysight B2902A source meter using a two-probe method. The optical stimuli are determined using a UV LED source of 375 nm connected to a second channel in the source meter. The LED is kept at a distance of 1.5 cm throughout the measurements. The ultrasonicated h-BN layers show photoluminescence, as shown in Fig. 1(e). Light emission from various defects can be discerned at various wavelengths and follows our previous findings.⁴³ The maximum intensity of PL is found to be about 374 nm, again prompting us to choose an optical light of 375 nm for optical pulsing. A typical optical pulse applied to the device is shown in Fig. 1(f). The UV light is kept ON for 4 s and then switched off while monitoring the postsynaptic current. Initially, the photocurrent increases quickly until the light is ON, and once the light pulse is switched off, the photocurrent decays. As proposed by Ebbinghaus, the current decay has two parts; the first part contains a faster decay followed by a slower reduction in the post-synaptic current. We monitor the decay of postsynaptic current after the removal of the optical pulse. The postsynaptic current increase/decrease when the UV light pulse is applied to the device is equivalent to the enhancement/reduction in memory of the human brain. When the pulse is removed, the conductance starts to decay, similar to forgetting in the human brain. In the following, the decay of normalized conductance is fitted with the Ebbinghaus curve for further analysis.

To check the dependence of the postsynaptic current and its decay, we altered the time of exposure of the device to the UV light. As shown in Fig. 2(a), the decay is monitored for a pulse duration of 4.2 s. Using the Ebbinghaus decay function, we get the time constant τ and the exponent β as 2.39 s and 0.415, respectively. Fitting the decay curves obtained for different pulse widths [as shown in Fig. 2(b)] shows that the characteristic time constant increases with exposure time while the exponent β remains almost constant. The values of the time constant and the exponent β indicate that the memristor based on h-BN can emulate the memory loss in the human brain. The cumulative effect of the increasing pulse width on the time constant τ and β value is plotted in Fig. 2(c). The characteristic decay time(τ) increased from 366 ms to 4.72 s for a 20 s long pulse. The time constant increases linearly, indicating that the memory state of the optical synapse is stable and can be linearly tunable. The conductance enhancement due to a larger pulse width has a different dependence on the exposure time. The change in the conductance linearly increases for short pulses but saturates for larger pulse widths, as shown in Fig. 2(c). The increment in ΔG_max with exposure time shown in Fig. 2(d) supports the short-term memory (STM) to long-term memory (LTM) transition. Our results follow the forgetting curve of the human brain, indicating that the h-BN-based optical synapse is a very good candidate to emulate biological synaptic functions.⁴⁴ 

In the case of the human brain, short-term memory (STM) and long-term memory (LTM) are two of the most important aspects of memory. The information is stored in the hippocampus as STM due to weak synaptic plasticity. This memory is lost in a very short time duration. If repeated learning is performed, then a long-term memory (LTM) is created. LTM can last for a longer duration, and the memory is now shifted to the cerebral cortex. Thus, short-term memory, long-term memory, and transition from STM to LTM have been demonstrated in solid state memristor devices. In the case of optical synapse, this depends on the number of optical pulses and also the frequency of optical pulses used. We checked the STM and LTM in the h-BN-based optical synapse and the possible transition from STM to LTM by changing the number of pulses applied to the optical synapse. The effect of increasing the number of optical pulses applied to the synapse is shown in Fig. (3). For a given 20 100 ms pulses, the decay follows precisely the Ebbinghaus forgetting curve with a time constant τ = 1.4 s and β = 0.32. Performing such measurements for an increasing number of pulses shows us that the time constant τ is linearly dependent. The exponent β remains unchanged, indicating that the decay again follows the Ebbinghaus curve and emulates the forgetting curve of the human brain. The linearly dependent time constant indicates that the memory in the optical synapse is constant and can be controlled efficiently.

Another way to establish the STM to LTM conversion is by applying optical pulses at different frequencies. Here, we applied 25 UV light pulses of varying frequency from 1.35 to 2.16 Hz. The pulse width used is 420 ms at a pulse interval of 42 ms. As shown in Fig. 4(a), the postsynaptic current decays according to Ebbinghaus’s forgetting curve. The values of time constant τ and β were found to be 11.7 s and 0.63, respectively. Similar measurements are carried out with various frequencies of optical pulses, and the results are shown in Fig. 4(b). The linear dependence of the time constant τ again indicates that the memory state could be controlled efficiently in the optical synapse. The enhancement of ΔG in the conductance state shown in Fig. 4(c) also shows a linear dependence on the frequency, which shows that the memory state can be controlled with the frequency of optical pulses as well.

Finally, the measured postsynaptic current in our devices is in the sub-nanoampere range, indicating very low power consumption of the device during the operation. For example, when we set a 100 ms optical pulse, the change in the postsynaptic current is about 0.057 nA. This means a pulse energy consumption of 2.85 pJ by the synaptic device. This will be an ideal device system for low energy-consuming neuromorphic applications.

To summarize, a planar optoelectronic synaptic device based on multilayered low dimensional h-BN is fabricated with platinum electrodes separated by a 5 μm gap. The device shows excellent short-term synaptic memory, which can be transformed to a long-term memory by increasing the UV exposure time, number of pulses, and frequency of UV light pulses. The Ebbinghaus forgetting behavior is emulated on these devices. The forgetting curve is emulated by applying a single optical pulse of different widths, different numbers of pulses, and pulses at different frequencies. In all cases studied, the conductance decay is found to emulate the Ebbinghaus forgetting behavior in the human brain. This confirms that low dimensional h-BN-based optical synapses can be used to emulate the biological forgetting curves.

The authors would like to thank the Center for Functional Materials, Vellore Institute of Technology, for their support during this research work. A.S. acknowledges the institutional fellowship during her Ph.D. work. R.T. would like to acknowledge the Core Research Grant (Grant No. CRG/2022/005093) from the Science and Engineering Board (SERB), India.