Bipolar Plasticity in Synaptic Transistors: Utilizing HfSe₂ Channel with Direct-Contact HfO₂ Gate Dielectrics

Abstract

The investigation of dual-mode synaptic plasticity was conducted in thin-film transistors (TFTs) featuring an HfSe₂ channel, coupled with an oxygen-deficient (OD)-HfO₂ layer structure. In these transistors, the application of negative gate pulses resulted in a notable increase in the post-synaptic current, while positive pulses led to a decrease. This distinctive response can be attributed to the dynamic interplay of charge interactions, significantly influenced by the ferroelectric characteristics of the OD-HfO₂ layer. The findings from this study highlight the capability of this particular TFT configuration in closely mirroring the intricate functionalities of biological neurons, paving the way for advancements in bio-inspired computing technologies.

1. Introduction

Recently, the exploration into emulating the brain’s complex functionality with electronic devices has sparked considerable interest, notably in the realm of neuromorphic engineering [1,2,3,4,5]. At the heart of neuronal functionality is the concept of synaptic plasticity, encompassing both potentiation and depression [6]. Despite the proliferation of diverse electronic devices designed to emulate synaptic plasticity, achieving a faithful replication of both plasticity aspects within a single device remains a formidable challenge, primarily due to the fundamental differences between biological synapses and their electronic counterparts [7]. Consider, for example, the brain’s ‘profit-and-avoid’ characteristic, these properties represent subtle neural responses to external stimuli, intricately dependent on the selective filtering mechanisms within neurons, enabling them to generate two distinct types of responses [8]. In essence, this ability enables neurons to generate two distinct responses. However, electronic devices often simulate a fixed synaptic function with limited single adaptability [9], thereby constraining their effectiveness in accurately mimicking complex biological synaptic processes, such as the brain’s ‘profit-and-avoid’ characteristic. This gap highlights the need for innovative approaches to more accurately model the brain’s complex synaptic processes.

In response to this formidable challenge, researchers have turned their attention to HfO₂ material, which is known for its high dielectric constant [10]. HfO₂ has become a leading material in the semiconductor industry, owing to its seamless integration with silicon-based industrial processes [11,12,13]. Recent findings have revealed that extremely thin layers of HfO₂ demonstrate ferroelectric properties, which has generated considerable interest due to their potential in influencing synaptic behavior [14,15,16,17]. This interest stems from the distinct response of ferroelectric materials to electrical stimuli, which opens up new avenues for application. While there are various ferroelectric materials, including perovskites [18], their incompatibility with the semiconductor industry’s standard processing methods precludes their use, especially in the fabrication of ultrathin layers, which is crucial for maintaining reliability and precision. In contrast, silicon remains a fundamental element in semiconductor technology, largely attributed to its 1.1 eV band gap enabling operations at low voltage [19]. The function of SiO₂ as a high-quality and inherent insulator has played a crucial role in maintaining silicon’s dominance in the industry for over fifty years. Nevertheless, in the past ten years, there has been a transition from using SiO₂ to HfO₂-based high-k insulators in silicon electronics [20]. Despite the benefits of HfO₂, it is not a native oxide of the silicon substrate, leading to various interface challenges and a reduction in ferroelectric effects due to inadequate polarization screening. This situation poses a compelling question: if silicon derives significant benefits from its native SiO₂ insulator, could HfO₂, a well-known high-k dielectric, inherently complement other semiconductors? Layered two-dimensional semiconductors, like HfSe₂, are significant in this scenario due to their band gaps, which vary from 0.9 to 1.3 eV. These band gaps cover a range from thicker bulk layers to thinner monolayers. HfSe₂ is considered a promising candidate for technological applications due to its compatible bandgap and its ability to work well with HfO₂’s dielectric properties. To mimic the ‘SiO₂/Si’ structure and explore new frontiers in semiconductor technology, a ferroelectric HfO₂/HfSe₂ stack has been adopted. This innovative approach potentially marks the beginning of a new era in semiconductor technology, with HfSe₂ playing a crucial role in the architecture of transistors, promising advancements in both functionality and efficiency.

In this context, devices containing HfSe₂/HfO₂ layers exhibited a dual characteristic in synaptic plasticity under the influence of various pulses [21]. Specifically, it exhibited an enhanced response to negative gate pulses and a subdued response to positive ones, mirroring the selective responsiveness of neurons to favorable or adverse stimuli. In these structures, the magnitude and polarity of voltage pulses emerge as pivotal factors. When negative-polarity pulses are applied, electrons are liberated within the HfSe₂ channel, resulting in an augmentation of channel conductivity. Simultaneously, polarization effects within the HfO₂ dielectric layer start to become apparent. The free electrons move rapidly due to the applied voltage, increasing the conductivity of the device. Simultaneously, when positive-polarity pulses are introduced, more charges become confined at the interface or permeate the HfO₂ layer. Consequently, the introduction of positive-polarity pulses during the pulse training process leads to a gradual reduction in response current. Throughout this process, the charge dynamics, influenced by the polarization behavior, mimic the release and reuptake of neurotransmitters. This behavior demonstrates the capacity of these electronic devices to accurately replicate biological synapses, effectively adjusting synaptic plasticity by facilitating both potentiation and depression responses. Furthermore, even when pulses are administered subsequent to the preceding pulse, polarization effects persist. The convergence of these two identical dynamic processes results in an augmentation of current, yielding the effect of increased current, akin to the behavior observed in biological synapses known as paired-pulse facilitation (PPF). Negative-polarity pulses induce electron release and amplify channel conductivity, while positive-polarity pulses prompt charge accumulation and a decline in conductivity. This bidirectional response underscores the potential of electronic devices to faithfully replicate the behavior of biological synapses. The dual nature of synaptic plasticity exhibited by these devices holds immense promise for emulating complex synaptic processes within biological systems.

2. Results and Discussion

Figure 1a presents a representation of a synaptic junction, a part of neural communication. It captures the process of neurotransmitter release from the presynaptic neuron and their reception by the postsynaptic neuron. The synaptic efficacy, the strength of the synaptic signal, is not static; it can be regulated by various factors such as the presence of an activating signal and the dynamic flux of neurotransmitters themselves. This regulation ensures that neural communication is not only robust but also plastic, capable of adapting to different physiological conditions and demands. the gate electrode’s voltage pulse is conceptualized as the activating signal, with the consequent current between the source and drain electrodes serving as a proxy for synaptic efficacy. That is, in a fashion analogous to this biological process, the voltage pulse applied to the gate electrode of a transistor can be envisioned as an activating signal. When a voltage pulse is applied, it modulates the conductivity between the source and drain electrodes of the transistor, akin to the way an action potential facilitates the release of neurotransmitters at the synaptic junction. The flow of current that results from this modulation serves as an electrical counterpart to the synaptic efficacy observed in biological systems.

Figure 1b shows the specific construction of a field-effect transistor (FET) in order to mimic the functions of a synaptic junction. At its core lies the channel made of HfSe₂ layers, and close to this channel, there is a layer of HfO₂, which is oxygen-deficient, acting as the dielectric material. The presence of oxygen vacancies in HfO₂ is critical for the ferroelectrical characteristic which further affects the overall behavior of the FET.

The transfer characteristics of the transistor, which relate the drain-source current (I_DS) to the drain-source voltage (V_DS), are graphically represented in Figure 1c. This graph shows the I_DS response of the transistor to varying gate-source voltages (V_GS) in increments of 1 volt, ranging from 1 to 6 volts. Each curve corresponds to a specific V_DS, and as V_DS increases, there is a noticeable increase in I_DS for a given V_DS, demonstrating the transistor’s ability to modulate current flow in response to changes in gate voltage, much like a neuron’s response to different levels of stimuli.

The transistor demonstrates notable operational characteristics, including a threshold voltage of approximately −1.3 V, boasting an ON/OFF ratio exceeding 2.3 × 10⁶, a commendable field-effect mobility of approximately 1.1 × 10 cm²/V·s, and an impressive subthreshold swing measuring merely 0.12 V/decade. Figure 1d provides a view of the threshold behavior of the transistor, detailing the subthreshold and above-threshold conduction regions. The graph plots I_DS on a logarithmic scale against V_GS, illustrating the sharp increase in current as the gate voltage crosses a critical threshold—a behavior that is reminiscent of the all-or-nothing response of a neuron when it fires an action potential. The red and black data points likely represent different measurements sequence that show the consistency of the transistor’s behavior in response to varying gate voltages. Within Figure 1d, we present the transfer characteristic curve, which results from a controlled 5 V bias applied across the source and drain electrodes. Interestingly, we observe a subtle hysteresis phenomenon during the gate voltage cycle, spanning from −3 V to 5 V and returning. This behavior hints at the potential occurrence of charge trapping events, either at interface junctions or within the gate dielectric. The presence of hysteresis is further corroborated by the directional arrows noted during the voltage sweep, providing empirical support for its existence and lending weight to the charge trapping [22,23,24].

Figure 2 provides a detailed visualization of the behavioral patterns of a transistor when subjected to a carefully designed sequence of bipolar voltage pulses. Figure 2a demonstrates the intentional use of a sequential pulse train applied to the transistor’s gate. Each pulse is precisely timed with a periodicity of 20 ms, has a fixed amplitude of 1 V, and is delivered with a pulse width of 20 ms. The accurate and regular stimulus is crucial for manipulating the gate of the transistor and imitating synaptic behavior.

Figure 2b captures the resultant current traversing the channel at a drain-source voltage (V_DS) of 4 V. What is particularly striking in this depiction is the transistor’s dynamic current response, which shows a significant potentiation effect in reaction to the negative pulse trains, whereas a clear depression effect is evident with the imposition of positive pulse trains. The contrasting directions of current modulation—increasing with negative pulses and decreasing with positive ones—emphasize the transistor’s capability for bidirectional programming. This critical functionality is instrumental in adjusting synaptic weight and closely replicates the bipolar pulse stimuli effect observed in biological synaptic interactions.

The observed phenomena can be attributed to charge trapping and detrapping processes occurring within the structure of the transistor, which further confirms the previously demonstrated trap activity. The initial application of a negative pulse triggers the release of trapped charges. With each subsequent pulse, this release effect is amplified, resulting in a substantial increase in the channel’s conductance when a negative pulse train is applied. Conversely, the application of positive pulses achieves the opposite effect, diminishing the channel’s conductance. This charge modulation, which dynamically influences synaptic weight, parallels the neurotransmitter dynamics seen in biological synapses, thus heralding a new avenue for semiconductor applications that could mirror the adaptive modulation of synaptic efficacy observed in natural biological processes. What is particularly striking in this depiction is the transistor’s dynamic current response, which shows a significant potentiation effect in reaction to the negative pulse trains, whereas a clear depression effect is evident with the imposition of positive pulse trains. The contrasting directions of current modulation—increasing with negative pulses and decreasing with positive ones—emphasize the transistor’s capability for bidirectional programming. This critical functionality is instrumental in adjusting synaptic weight and closely replicates the bipolar pulse stimuli effect observed in biological synaptic interactions.

The observed phenomena are attributable to charge trapping and detrapping, corroborated by trap evidence presented in Figure 1d. The initial application of a negative pulse precipitates charge release, with subsequent pulses exacerbating this effect, thereby amplifying the channel’s conductance under source/drain (S/D) bias via negative pulse train stimulation, and inversely with positive pulses. Analogous to synaptic neurotransmitter dynamics depicted in Figure 1a, this charge modulation process dynamically alters synaptic weight. This mechanism indicates possible applications that resemble biological synapses, where the effectiveness of synapses is dynamically adjusted by bipolar spike protocols. Furthermore, the enduring characteristic of the response current, which is hypothesized by the use of low-intensity reading pulses (0.2 V, 20 ms) after the pulse train has stopped, is recorded in Figure 2c,d. The response current exhibits a progressive increase or decrease, contingent on the polarity of the applied read pulse. This sustained response, persisting beyond the stimulus signal’s duration, suggests an underlying influence beyond mere charge dynamics, potentially linked to the documented ferroelectric properties of HfO₂. Furthermore, Figure 2c,d capture the enduring behavior of the response current, which is meticulously tracked through the use of small amplitude reading pulses of 0.2 V for a duration of 20 ms, administered following the termination of the pulse train. The response current’s behavior, as documented, exhibits a consistent and progressive change in magnitude, corresponding directly to the polarity of the read pulse applied. This persistent change, which outlasts the stimulus signal itself, hints at an extended influence that transcends simple charge movement and may be intimately linked to the ferroelectric characteristics of materials like HfO₂, which have been previously noted in other research.

To discern the influence of pulse parameters on the response current in neuromorphic transistors, an exploration was conducted by varying one of the three defining parameters—amplitude, interval, and width—while holding the others constant [25]. Alterations to the other parameters of the pulse sequence, interval or width, did not yield significant shifts in response current. This absence of variation suggests that carrier release and capture are contingent upon reaching a critical energy threshold, principally determined by the voltage amplitude. This finding further corroborates that the pulse conditions applied in this study were sufficient to activate the traps effectively while leaving deeper-level traps unaltered. This observation suggests that the act of carrier release and subsequent capture within the transistor’s framework relies on surpassing a pivotal energy boundary, which is predominantly dictated by the voltage amplitude. This unresponsiveness to changes in pulse interval or width thus reinforces the assertion that the pulse conditions selected for this study were optimally chosen to activate the charge traps with high efficacy, while having a negligible effect on deeper-level traps that may not be as readily influenced by the chosen stimulus parameters.

Figure 3a illustrates the synaptic-like behavior of a neuromorphic device in response to a temporal sequence of pulse trains at varying magnitudes. The primary graph reveals the device’s response current over an extended time period, with the application of pulse trains ranging from −2 V to +2 V. Each color represents the response to a specific voltage magnitude, indicating how the current either increases or decreases in response to the applied voltage over time. Several key observations can be made: (i) Magnitude-Dependent Responses: There is a clear magnitude-dependent response in the device, with higher absolute voltages eliciting larger currents. This suggests that the device’s conductance is sensitive to the voltage magnitude, an important characteristic for emulating the plasticity of biological synapses. (ii) Temporal Dynamics: The response currents display a temporal dependence, highlighting the device’s ability to retain a memory of the voltage stimulus. This temporal aspect of the response is critical for mimicking the time-dependent processes found in biological synaptic behavior. (iii) Polarity Sensitivity: The device demonstrates a symmetric response to the polarity of the applied voltages, with positive voltages inducing positive currents and negative voltages inducing negative currents. This symmetry may reflect an intrinsic property of the device’s operational mechanism, which is essential for replicating the bidirectional modulation of synaptic strength. (iv) Detailed Observation from Inset: The inset provides a magnified view of the response current at a finer scale (in microamperes) and shows the response after read pulses at various time intervals. This could be instrumental in detecting subtle changes after prolonged sequences, akin to the regulatory mechanisms in biological synapses. (v) Time-Dependent Changes: The response current changes over time suggest a certain degree of potentiation or depression, akin to the learning and memory functions of biological synapses, which may be influenced by the device’s charge trapping dynamics.

The graph depicts the potential of a neuromorphic device to exhibit synaptic plasticity, with a current response that is highly sensitive to the amplitude and temporal characteristics of voltage pulses. Pulse parameters—amplitude, duration, and periodicity—are pivotal in modulating the neuromorphic transistor’s response current. Our study focused on the amplitude’s influence while maintaining consistent pulse duration and periodicity. Incrementing the amplitude of negative/positive pulses resulted in an augmented/diminished synaptic response current (refer to Figure 3a), suggesting enhanced efficacy in the charge’s release/trapping processes. The augmentation in current under a negative pulse and its reduction under a positive pulse reflect the amplitude’s role in controlling the charge dynamics. The supplementary inset of Figure 3a presents the response current continuity post pulse train cessation, utilizing a 0.2 V reading voltage. Notably, this current persists in its incremental/decremental trend under the accumulation of negative (positive) bias voltage, reinforcing the assertion that longer intervals of pulse do not abate the synaptic response’s progressive trajectory. This observation alludes to an intrinsic polar electric field’s sustained influence, validated through P-V measurements that substantiate the ferroelectric nature of the examined HfO₂ films (as shown in Figure 4b) [26]. Figure 4b illustrates the ferroelectric characteristics of HfO₂ thin films at various thicknesses (14 nm, 20 nm, and 30 nm), as shown by their polarization–voltage (P-V) loops. The P-V loops indicate that the ferroelectric properties of the films are not significantly dependent on the thickness, as all three thicknesses show similar hysteresis loops, which is characteristic of ferroelectric behavior. Notably, even at a reduced thickness of 14 nm, the film exhibits distinct ferroelectricity, which is often not as pronounced in other ferroelectric materials at such thin dimensions.

HfSe₂ is widely recognized as an n-type semiconductor. Under the influence of a ferroelectric field and external bias, oxygen vacancies in the HfSe₂ channel readily undergo ionization, and the energy barrier for neutralizing these ionized vacancies is remarkably low. Electrons within the HfSe₂ channel can traverse the HfSe₂/HfO₂ interface and subsequently become trapped within high value of pulse sequence. The observed response exhibits two primary behaviors: (i) The conductivity of HfSe₂ strongly depends on its charge carriers. Under negative polar pulse, it induces the liberation of electrons, thereby conductivity promoting within channel. Simultaneously, polarization initiates within the HfO₂ dielectric. The move of these liberated electrons, driven by the voltage (assisted by the polarization within the ferroelectric layer), results in enhanced device conductivity. The above is shown in Figure 4a,b. (ii) Following the cessation of a pulse, the restoration of negative charges is expected. Nonetheless, the inherent polarization of the OD-HfO₂ layer obstructs this recuperative mechanism. As a consequence of this dynamic interaction, it can be observed that, during the intervals between pulse trains, the response current to the modest read pulse (0.2 V) exhibits ongoing fluctuations, which become more pronounced with the introduction of more sparsely distributed read pulses. This phenomenon is graphically illustrated in Figure 4c. (iii) After applying a pulse following another one, polarization persists, and the combination of these identical dynamic processes enhances conduction, leading to an increase in the current. The visual representation of this phenomenon can be observed in Figure 4d. It is worth noting that the postsynaptic current elicited by the second pulse surpasses that of the initial pulse, resembling the behavior observed in paired-pulse facilitation (PPF) within biological synapses. [27].

Paired-pulse facilitation (PPF) is a form of synaptic plasticity characterized by an augmentation in postsynaptic responses when the second spike closely follows the preceding one, reflecting a potentiation phenomenon [28]. Recent reports have extensively covered the ferroelectric characteristics of HfO₂. The emergence of ferroelectricity in HfO₂ films is attributed to the existence of a metastable and non-centrosymmetric orthorhombic phase, characterized by the space group Pca21 [29]. The enhancement and control of ferroelectricity can be achieved by introducing various dopants and applying contact stress. However, when a positive gate voltage is applied, a greater number of charges are either trapped at the interface or pass through the HfO₂ layer. Consequently, the response gradually declines as the pulse trains added. The observation of charge trapping and the retention of polarization post pulse application in both sequences suggest a memory effect within the device, critical for neuromorphic applications. The ability to modulate this effect with the polarity of the applied pulses could be harnessed for simulating synaptic plasticity, thereby emulating the fundamental properties of biological synapses in artificial devices.

3. Materials and Methods

In the experimental setup, substrates and bottom electrodes made of n+ silicon were utilized. The initial cleaning process involved submerging the silicon wafers in a dilute hydrofluoric acid solution to effectively strip away any surface oxides and impurities, maintaining a ratio of 1% HF to water. Following this, a thin film of HfO₂, measuring 5 nm in thickness, was deposited onto the silicon substrates. The deposition process under investigation herein was conducted through the utilization of radio frequency (rf) magnetron sputtering as the primary deposition technique. This process entailed the employment of a high-purity Hafnium target, with a purity exceeding 99.99%, which played a pivotal role in ensuring the quality and purity of the deposited material. The deposition took place within a rigorously controlled environment characterized by a blend of Argon and Oxygen gases, affording precise control over the deposition conditions. In order to maintain an atmosphere conducive to the formation of oxygen-deficient oxide layers, the oxygen bias was meticulously regulated to attain a level of 8 × 10⁻⁵ Torr, while the Argon bias was set at 2.1 × 10⁻³ Torr. This fine-tuning of the gas pressures was pivotal in achieving the desired low oxygen environment, which is known to be optimal for the deposition of such layers. The material of interest, HfSe₂, was sourced from HQ Graphene in the form of sheet-like crystals, boasting facets measuring several millimeters in size. The synthesis of these HfSe₂ layers was carried out through the application of chemical vapor transport (CVT) methodology. Elemental Hafnium and Selenium precursor powders were judiciously employed in the synthesis process, with utmost precision and control, ultimately resulting in the successful production of the desired material. This synthesis approach and the choice of precursor materials were instrumental in ensuring the high quality and purity of the synthesized HfSe₂ layers, as demanded by the scientific investigation at hand. These HfSe₂ layers were then mechanically exfoliated and carefully transferred onto the prepared HfO₂ layer on the silicon substrate. This transfer was executed with precision, ensuring optimal interface quality between the HfSe₂ and HfO₂ layers. After the successful integration of the HfSe₂ layer onto the HfO₂, the next step involved the fabrication of electrodes. The experimental procedure encompassed the deposition of two distinct layers of metallic films, specifically 200 nm of gold (Au) and 50 nm of titanium (Ti), functioning as the source and drain electrodes. This deposition process was meticulously executed through electron beam evaporation, ensuring precision and control over the film thickness and material properties. Subsequently, these deposited electrodes underwent a patterning process to achieve specific dimensions. The resulting electrodes were designed to possess lengths of 60 μm and widths measuring 1500 μm, in accordance with the experimental requirements and design specifications. A forming gas anneal (FGA) was applied post-electrode deposition to improve the metal-semiconductor contact. The final stage in the device fabrication was the thermal annealing process. A post-deposition annealing at 300 °C in a nitrogen atmosphere for 120 s was performed to optimize the device properties. The fully assembled device, which included the HfSe² and HfO² layers in conjunction with the meticulously patterned source and drain electrodes, underwent a series of comprehensive electrical characterizations. These characterizations were aimed at assessing its electronic properties and performance. To carry out these assessments, precise current–voltage (I–V) characteristic measurements were conducted. The measurements were conducted using state-of-the-art instrumentation, including a Keithley 4200 SCS system. Additionally, an Agilent B2900 Precision Source/Measure Unit (SMU) was employed to provide precise control over the electrical parameters during the measurements.

4. Conclusions

In this study, we developed thin-film transistors (TFTs) utilizing a HfSe₂ channel. These advanced TFTs demonstrated an augmentation in the post-synaptic current upon the application of negative pulse to bottom gate. Conversely, positive bias resulted in a diminution of the post-synaptic current. This phenomenon is attributable to the modulation of carrier concentrations within the HfSe₂ layer, with the ferroelectric tendencies of the hafnium oxide layer serving as a facilitator. Notably, the biphasic response behavior of these transistors underlines their immense potential in emulating biological synapses, paving the way for bioinspired neuromorphic applications.