The computing capability of classical digital computers, based on Complementary Metal Oxide Semiconductor (CMOS) transistors, has advanced considerably in the past decades, mainly due to the shrinking down of transistor’s dimensions, as predicted by Moore’s law (Moore, 1965). The advent of the Artificial Intelligence (AI) has imposed critical requirements in terms of energy efficiency and processing speed, to address ambitious problems such as speech and image recognition (Gokmen and Vlasov, 2016). Conventional von Neuman architectures face two main challenges: first, Moore’s law is slowing down (due to rising fabrication cost and physical limitations), second, their performance is limited by the data transfer between the processor and the memory (Wong and Salahuddin, 2015). Brain-inspired neuromorphic architectures, allowing to perform computing at the site where data is stored, hence in-memory, are promising candidates to overcome this issue (Poon and Zhou, 2011). Such architectures consist of a collection of artificial neurons interconnected by plastic synapses in a crossbar topology which allows to efficiently perform the multiply and accumulate operation (Gokmen and Vlasov, 2016), a key computing task in neural networks (Kim et al., 2017; Yu, 2018). To imitate the biological synaptic plasticity, an analog programming capability of these synapses is required to define the synaptic weight. To achieve densely integrated neuromorphic circuits, both the material and the processes of the synaptic devices are required to be compatible with modern CMOS technology. Several technology implementations and physical phenomena, such as Phase-Change Memory (PCM) (Lacaita, 2006; Raoux et al., 2010; Boybat et al., 2018), filamentary-based Resistive Random Access Memory (RRAM) (Baek et al., 2004; Lee et al., 2008; Waser et al., 2009) and Electro-Chemical Memory (ECRAM) (Fuller et al., 2017; Kim et al., 2019; Tang et al., 2019), can lead to synaptic behavior, but they all rely on structural modification of the active materials involved. The recent discovery of the ferroelectric properties in hafnia composites (Böscke et al., 2011), a material already present in CMOS technology, has attracted further scientific interest in the field of neuromorphic hardware based on ferroelectrics. Three main classes of devices exploiting ferroelectricity for synaptic as well as neuronal functionalities were demonstrated in the past: the two-terminal Ferroelectric Tunneling Junctions (FTJs) (Ambriz-Vargas et al., 2017; Tian and Toriumi, 2017; Chen et al., 2018a; Goh and Jeon, 2018; Yu et al., 2021), the three-terminal Ferroelectric Field-Effect Transistors (FeFETs) (Mulaosmanovic et al., 2017; Sharma et al., 2017; Krivokapic et al., 2018; Zeng et al., 2018; Mo et al., 2019) and the two-terminal Ferroelectric Photovoltaic (FePv) synapses (Cheng et al., 2020; Cui et al., 2021). Although, both FTJs and FeFETs have been extensively investigated recently, showing large dynamic ranges, low energy dissipation, and synaptic functions including short and long term plasticity as well as Spike-Timing-Dependent Plasticity (STDP) (Nishitani et al., 2012; Boyn et al., 2017; Chen et al., 2018a; Guo et al., 2018; Majumdar et al., 2019; Li et al., 2020) the FePv devices, based on the polarization control of the photovoltaic behavior that exploit the photoresponsivity as synaptic weight, were used for binary data storage (Guo et al., 2013) and recently as prototype synapse (Cheng et al., 2020). While in state of the art hafnia-based two-terminal synaptic weights, the small current flowing through the ferroelectric layer limits their scalability (Begon-Lours et al., 2021), three terminal devices have the advantage of separating the write process (through the high impedence gate) and the read process (through the channel). Hafnia-based FeFET devices exploiting Si as channel material and implemented on the Front End Of Line (FEOL) were demonstrated as artificial neurons (Mulaosmanovic et al., 2018). However, the FEOL integration imposes constraints on the device footprint, and limits the design flexibility. The Back End Of Line (BEOL) integration is advantageous, by allowing for a larger device area, which in turn leads to a larger number of ferroelectric domains and, hence, an improved analog (multi-level) behavior. Planar state of the art BEOL three-terminal synaptic devices based on HfZrO₄ (HZO) and utilizing a tungsten oxide (WO_x) channel, were realized in the past (Halter et al., 2020). However, in the last decade, the tri-gate technology (Lawrence and RUBIA, 2015) has replaced the planar one and allowed further CMOS transistor scaling. In this architecture, the gate surrounds the channel on three sides, creating a multigate device known as FinFET, with better gate-channel control and a smaller footprint with respect to a planar technology. In this work, we report on a scaled tri-gate FeFET (FinFeFET) having an overall footprint scaled down to four orders of magnitude with respect to (Halter et al., 2020). Being a Junction-Less Transistor (JLT) (Colinge et al., 2010; Colinge et al., 2011), no high-temperature source and drain implantation and annealing processes are required during the fabrication. The synaptic behavior is achieved through the partial polarization switching in HZO, which is used to electrostatically deplete or accumulate free carriers in the WO_x fins. We demonstrate the scaling of the ferroelectric technology down to device having 0.002 µm² area, and study the impact of the layout on the channel resistance, the influence of the fin’s geometry on the dynamic range, the retention, the analog behavior as well as the continuous and linear synaptic weight modulation. Moreover, both the process and the materials exploited are compatible with CMOS technology, the proposed synaptic element is promising for large-scale and densely integrated neuromorphic hardware based on ferroelectrics.

3D FeFET devices based on a W/TiN/HZO gate stack and 30 nm high WO_x fins were designed and fabricated using a process BEOL compatible, not exceeding 400°C. To investigate the effect of the layout on the device performances, several geometries of FinFeFETs were processed to find out the best trade-off in terms of fin’s length, width and number. Fins of 4 nm, 8 and 10 nm width were explored, and for each of them, two different lengths, 200 and 500 nm, respectively, and configurations with 1, 5, 10, 20, and 40 parallel fins were fabricated. The substoichiometric and amorphous WO_x channel, deposited by a Plasma-Enhanced Atomic Layer Deposition (PEALD) process at 375°C, was first crystallized and oxidized by annealing in an oxygen atmosphere, and then structured into fins. The source and drain contacts were deposited on the WO_x channel through lift-off. Then the TiN/HZO stack was grown, and the ferroelectric crystallization of the latter was performed using a millisecond flash lamp anneal at 375°C. The device was encapsulated by a 5 nm of Al₂O₃ and a 100 nm of SiO₂ passivation layers. Contact pads were formed on top of the passivation layers and routed through openings to source, drain and gate. The detailed processing steps can be found in the Sub-section 4.1. In Figure 1A the result of the fabrication process after the FinFeFET contact lift-off step was imaged by a Scanning Electron Microscope (SEM). SEM analysis of the fins revealed that the targeted widths of 4 and 8 nm both resulted in an approximately 10 nm wide fin after the transfer of the design from the resist to the WO_x. This is the result of cross exposure of dense structures close to the resists resolution limit. The materials properties were characterized by Grazing-Incidence X-Rays Diffraction (GIXRD). Figure 1B shows the GIXRD (ω = 0.44°) performed after HZO crystallization by ms-flash lamp annealing: the peak at 2θ = 30.8° corresponds to the overlap between the (111) peak of the orthorhombic (ferroelectric) phase and the (011) peak of the tetragonal phases of HZO (Park et al., 2013). As a consequence of the low temperature crystallization technique (O’Connor et al., 2018), no monoclinic HZO phase (peaks at 28.2° and 31.8°) (Materlik et al., 2015a) is observed in our sample. The additional peaks at 28.6°, 33.6° and 34.5° can be attributed to (111), (202) and (220) Miller indices of the tetragonal P42₁m phase of WO₃ (Jain et al., 2013), respectively. The peak at ≃ 36° is a combination of the multiple peaks from the orthorhombic and tetragonal phases of HZO. Figure 1C shows the two cross-section illustrations of the FinFeFET and its relative process flow. The resistive switching of HZO—WO_x FinFeFETs was investigated through electrical characterization. Oxygen vacancies confer n-type semiconducting properties to sub-stoichiometric WO_x<3 (Salje and Güttler, 1984). When the HZO ferroelectric remanent polarization points toward the interface with WO_x, free carriers accumulate at the interface to screen the polarization charges in HZO, thus the channel resistance R _SD decreases, and the memristor is in its Low Resistive State (LRS). By contrary, when the remanent polarization points toward the TiN interface, carrier depletion occurs in tungsten oxide at the interface with HZO, causing an increase of the channel resistance R _SD and resulting in a High Resistive State (HRS). The schematic energy band diagrams at the equilibrium of the TiN/HZO/WO_x stack, both in depletion (HRS) and accumulation (LRS) states, are shown in Figure 1D.

Polarization charges are screened in the HRS state and it is possible to define a screening length (depletion width) x _d , representing the thickness of the channel where the resistance is modulated. By decreasing the carrier density n W O x , the depletion width x _d is increased (Davis, 1973). Considering the FinFeFET devices with a single fin, the overall channel resistance R _SD can be thought as the resistance of two channels in parallel: one of thickness x _d on the outside of the fin, in which the resistivity is modulated upon polarization switching, and a bulky one which extends in the core of the fin, with a constant resistivity (Bégon-Lours et al., 2018; Halter et al., 2020). In multi-fin FeFETs, the equivalent channel resistance R _SD can be looked at as multiple single-fin resistances in parallel. For neuromorphic applications a large dynamic range (HRS/LRS), as well as multiple (analog) levels of the channel resistance, an absolute resistance in the tens of megohm range, good retention properties, a low device-to-device and cycle-to-cycle variation, a linear and symmetric weight-update rule, and a low power consumption are important characteristics of ideal memristors (Yu, 2018). The exact requirements vary depending on the application and from one implementation to the other. HZO is polycrystalline and the lateral domain size of HZO films prepared in the same conditions (O’Connor et al., 2018; Halter et al., 2020) is found to be ≃ 10 nm: a single fin is interfaced with several hundreds of domains. Switching only a subset of them allows the analog modulation of the channel resistance. To quantify the range of the pristine resistance from device to device, DC-electrical characterization was performed. A voltage sweep back and forth between −0.21 and 0.21 V, with a step of 20 mV, was applied between the source and the drain, keeping the gate floating. The electrical transport in the pristine state of FinFeFETs having different number of fins, is reported in Figure 2A. The current density J is calculated from the measured current I by J = I N f i n s ⋅ h f i n s ⋅ w f i n s , where N _fins , h _fins and w _fins are the number, the height and the width of the fins, respectively. For all the configurations, log(J) depends linearly on log(V) with a slope of 1, showing that the conduction in the channel is Ohmic. It depends on the carrier density by the relation (Sze and Ng, 2006): log J = log q ⋅ μ ⋅ n W O x L f i n s + log V (1) where q is the elementary charge, μ the electron mobility, and n W O x the carrier density in the channel. Averaging the intercepts showed in Figure 2A and exploiting Eq. 1, the extracted μ ⋅ n product in WO_x is 1.87×10²⁰ (cmVs)⁻¹. In absence of specific structures to perform Hall measurements, it was not possible to determine the electron mobility in the channel. However, using the value of the WO_x mobility ( μ W O x = 0.19 cm²/V s) extracted by (Halter et al., 2020), the estimated carrier concentration is n W O x = 9.84 × 10²⁰ cm⁻³, which is coherent with the n-type semiconducting properties of the sub-stoichiometric WO_x<3. The remanent polarization P _r of 10 nm HZO was measured on TiN/WO_x/HZO/TiN capacitors on a different sample by Positive-Up Negative-Down (PUND) measurements (see Sub-section 4.3) and found to be P _r− = −11.2 µC/cm² and P _r+ = 17.7 µC/cm² (see Figure 2B). The asymmetry between positive and negative remanent polarization (|P _r+| > |P _r−|) indicate partially switched domains due to incomplete screening by the depleted WO_x layer. This results in a depolarization field across HZO (Mehta et al., 1973). Using the polarization as total charge per unit area (Q _s ) we can define the electric field (E _interface ) induced by it at the HZO/WO_x interface by using Gauss law (Brotherton, 2013): P r − = Q s = − ϵ 0 ⋅ ϵ H Z O ⋅ E i n t e r f a c e (2) where ϵ _HZO is the permittivity of HZO, which for the ferroelectric phase is 29.1 (Materlik et al., 2015b). The depletion width x _d in WO_x with respect to the electric field caused by the polarization charges can be related by using Poisson’s equation Brotherton (2013): E i n t e r f a c e = q ⋅ n W O x ⋅ x d ϵ 0 ⋅ ϵ W O x (3) By combining Eqs 2, 3, the depletion width can be estimated as follows: x d = − ϵ 0 ⋅ ϵ W O x ⋅ P r − ϵ 0 ⋅ ϵ H Z O ⋅ q ⋅ n W O x = 4.5 n m (4) where ϵ W O x is the permittivity of WO_x, equal to 189 as measured by (Halter et al., 2020). A more precise determination of x _d is not possible due to the missing μ W O x of the fins, but this approximation confirms that the effect of the polarization on the depletion width is comparable to the channel dimension. The retention properties of FinFeFET have been studied, as shown in Figure 2C. First, the device was set in its LRS by applying a 500 µs pulse of V _write = 5 V. Then, the channel resistance R _SD was monitored and read at 200 mV every few minutes up to 1.5 × 10⁴ s. This was repeated for the HRS using a 500 µs pulse of V _write = −5.5 V, and the evolution of that state was monitored for the same time interval. By fitting a linear regression in the semi-log representation, a drift is observed towards lower values. Extrapolating the fit to 10 years, both the HRS and the LRS are still differentiable, but the dynamic range is reduced. Remarkably, both the states drifted towards lower values, indicating that this drift does not originate from the back-switching of ferroelectric domains caused by the depolarization field, but possibly from an oxygen exchange between the WO_x and HZO, which progressively reduces the channel. This is more pronounced after setting the HRS (oxygen drift from HZO to WO_x by negative write field). The pristine resistance R _SD of each device was extracted. A decreasing trend of R _SD as a function of the number of fins is reported for a representative FinFeFET configuration (Figure 2D). Since multiple-fin configurations are a convolution of the single-fin ones, and since for neuromorphic applications, the absolute resistance should be in the megohm range (Gokmen et al., 2016), further electrical characterization was performed only on single-fin devices, which achieved the targeted resistance and the smallest footprint. The resistive switching of a representative FinFeFET (Figure 2E) with a 10 nm wide and 200 nm long fin, was investigated by applying a DC voltage of varying amplitude V _write on the gate of the memristor, having the source and the drain as common reference, and reading the channel at V _read = 200 mV. A more detailed description of the writing and reading procedures can be found in the Sub-section 4.3. By sweeping V _write from −4 to 4 V, R _SD shows a hysteresis cycle from 1.5 MΩ to 2.4 MΩ (HRS/LRS ≃ 1.53). Set and reset operations occur with a positive and negative programming voltage on the gate, respectively. 30 single fin devices were characterized as previously explained, to measure the static dynamic range (Figure 2F). The variability in the dynamic range may be due to processing at the limit of our lithographic capabilities, as well as an inhomogeneous WO_x material.

Several pulsing schemes on HZO based FeFETs have been investigated in the past (Jerry et al., 2018). In this work, the scheme using pulses with varying amplitudes and constant width was used since it optimizes the number of accessible polarization states (Jerry et al., 2018). The analog nature of a representative FinFeFET (Figure 3A) having a 10 nm wide and 500 nm long fin, was explored by applying voltage pulses of varying amplitude V _write while keeping a fixed pulse duration of 5 µs. For the potentiation, V _write was increased from 1 to 5 V, and for the depression, decreased from −1 to −5.5 V, with 50 mV steps. A slightly higher voltage was used for the depression to compensate the built-in field in HZO. After each pulse, the channel resistance R _SD was measured at V _read = 200 mV, keeping the gate floating. The memristor showed a HRS of ≃ 1 MΩ and a LRS of ≃ 0.7 MΩ (HRS/LRS ≃ 1.4). With respect to the DC-electrical characterization, almost all the devices had a decrease in dynamic range. This may be explained considering the short programming pulses and that the dynamics of ferroelectric switching in polycrystalline HZO films follow the Merz law (Chanthbouala et al., 2012), (Paruch et al., 2006), hence the coercive field (E _c ) depends linearly on the logarithm of the writing time, as detailed for polycrystalline HZO devices in (Bégon-Lours et al., 2021). Another explanation could be an oxygen drift across the HZO—WO_x interface, a much slower process than ferroelectric switching, that would lead to an oxidation/reduction of the WO_x channel. The drift phenomena should be more pronounced in the DC potentiation and depression and thereby lead to the observed dynamic range dependence on the write signal length. The cycle to cycle variability was taken into account averaging all the potentiation and depression cycles (Figure 3B). By decreasing the range of V _write , the dynamic range is reduced. The number of the intermediate states is defined by the potentiation and depression step size, which can be further reduced to increase the resolution. The resistive states are not all differentiable, however the monotonic increasing and decreasing trends are desirable for online learning. The weight-update linearity was quantified by fitting the normalized weight update characteristics, by a function of the normalized pulse number, as proposed by (Chen et al., 2018b; Chen et al., 2018c): f x = 1 − exp − x A 1 − exp − 1 A (5) The parameter A was chosen by minimizing the root mean square error of the fitting. Values of A _LTP = 1.2 for the potentiation and A _LTD = −2.5 for the depression, respectively, were found (see Figure 3C).

Considering the ferroelectric synaptic weight dependence both on the pulse amplitude and duration, FinFeFETs are promising devices for Spike-Timing-Dependent Plasticity (STDP). However, STDP was not implemented with such devices in this work, since it requires tailored spike shapes as described by (Boyn et al., 2017) in ferroelectric perovskites and by (Max et al., 2020) in ferroelectric hafnia.

We developed a manufacturing process to allow the transfer and the scale-down of the FeFET planar technology into a multigate FinFeFET configuration. The fabrication process is compatible with the integration in the back end of line of CMOS technology and is using only abundant materials, making it suitable for large-scale integration. An Ohmic conduction in scaled WO_x fins, as well as good retention, analog states and an almost symmetric and linear potentiation and depression were obtained. Future work will focus on controlling the carrier concentration of WO_x fins, to further increase the resistance range and the dynamic range.