Hybrid 2D–CMOS microchips for memristive applications

1Our 2 cm × 2 cm silicon microchips have been designed by means of Synopsys software and fabricated in a 200 mm silicon wafer in an industrial clean room using a 180 nm CMOS technology node (Fig. 1a and Extended Data Fig. 1). The circuits fabricated in this study consist of 5 × 5 crossbar arrays of one-transistor-one-memristor cells (1T1M, Fig. 1b,c and Supplementary Fig. 1), although some standalone memristors and CMOS transistors were fabricated for reference (Supplementary Fig. 2). The microchips have been designed to integrate the memristors into the back-end-of-line (BEOL) interconnections; that is, they have been terminated at the latest metallization layer (fourth in our wafer) and have been left without passivation. Hence, silicon oxide naturally grows on the wafers when they are extracted from the industrial clean room (Fig. 1d), which can be easily etched away to expose the tungsten vias (Fig. 1e and Supplementary Fig. 3). Then, a roughly 18-layer-thick sheet of hexagonal boron nitride (h-BN) (that is, roughly 6 nm), grown on a Cu substrate by means of chemical vapour deposition (CVD), was transferred on the microchips (Fig. 1f) using a low-temperature process (Methods). Finally, the h-BN on the contact pads was etched, and top electrodes made of different materials (that is, Au–Ti, Au or Ag) were patterned and deposited on the h-BN to finalize the circuits (Fig. 1g).

2As the tungsten vias of the fourth metallization layer have a diameter of roughly 260 nm (Fig. 1h and Supplementary Fig. 3), the lateral size of the resulting h-BN memristors is, at most, 0.053 µm2. Figure 1h shows a high-angle annular dark-field cross-sectional scanning transmission electron microscope image of a 1T1M cell (with top Au–Ti electrode) in the crossbar array (Supplementary Figs. 4 and 5). The correct layered structure of the h-BN stack is confirmed before and after transfer by means of cross-sectional transmission electron microscopy (TEM) (inset in Fig. 1h and Extended Data Fig. 2). Nano-chemical analyses by means of electron energy loss spectroscopy demonstrate the correct composition of the h-BN sheet (Extended Data Fig. 3). The optical microscope images (Fig. 1c) reveal that the h-BN sheet does not crack during the transfer; this is an important advantage of using roughly 6-nm-thick 2D layered materials, and it increases the yield of the devices and circuits compared to counterparts using monolayer 2D materials.

3When we apply sequences of ramped voltage stresses (RVS) to several standalone 0.053 µm2 Au–Ti–h-BN–W structures, most (roughly 90%) of them show erratic current fluctuations and no resistive switching (RS) is observed (Fig. 2a); the currents do not reach linear regime (that is, dielectric breakdown) even if we apply 11 V. This is striking because most (more than 75%) h-BN devices with larger areas (25 µm2) show dielectric breakdown voltages (VDB) between 3 and 11 V followed by filamentary non-volatile bipolar RS. The reason should be the lower probability to find clusters of defects in small devices, which remarkably increases VDB (ref. ). Few (roughly 10%) 0.053 µm2 Au–Ti–h-BN–W structures show VDB between roughly 2.5 and 4 V followed by filamentary non-volatile bipolar RS (if a current limitation 1 mA or higher is applied, Supplementary Fig. 6). However, the endurance is only around 100 cycles, mainly due to the poor controllability of the current across the memristor and the overshoot during the dielectric breakdown.

4On the contrary, the CMOS transistor in the 1T1M cell can precisely control the current across the h-BN memristor and avoid the current overshoot, which results in outstanding performance. First, we obtain the output characteristic of one standalone CMOS transistor by applying a constant voltage to the gate (VG) and a RVS to the drain (VDS), and measuring the drain-to-source current (IDS); the CMOS transistor works correctly as expected (Fig. 2b). And second, we measure the 1T1M cell by applying RVS at the top Au–Ti electrode of the memristor while keeping the source terminal of the transistor grounded and simultaneously applying a constant VG. When a sequence of RVS is applied to the top electrode of the Au–Ti–h-BN–W structure using VG = 1.1 V, most 1T1M cells show non-volatile bipolar RS (Fig. 2c and Extended Data Fig. 4). The high state resistances (RHRS of roughly 200 MΩ and RLRS of roughly 200 KΩ)—beneficial to reduce power consumption—the non-linearity of the currents in both states and the progressive state transitions indicate that the RS is non-filamentary. However, we do see an activation process, as the first RVS slightly increases the conductance of the devices (that is, softly degrades the h-BN stack, blue line in Fig. 2c and Extended Data Fig. 4). In the first microchip that we fabricated, this stable non-filamentary bipolar RS regime was observed in 32 out of 40 cells (yield 80%), and in the last one it was observed in 25 out of 25 devices (yield 100%). On the contrary, standalone Au–Ti–h-BN–W structures of 0.053 µm2 and 1T1M cells without h-BN never showed this behaviour; this confirms that the RS is produced by the h-BN stack and that the CMOS transistor is key to control its soft degradation. Note that in the standalone Au–Ti–h-BN–W structures the current is limited using the semiconductor parameter analyser, which activation time is long (roughly 70 µs) and parasitic capacitance is high (roughly 300 pF, related to the cables); on the contrary, in the 1T1M cell the series transistor acts as an instantaneous current limitation (it cannot drive more current than that allowed by the size of its channel) and the parasitic capacitance is much lower (roughly 50 pF, internal connections in the microchip), which reduces the duration of the switching transient and undesired currents across the Au–Ti–h-BN–W structure. The values of RHRS and RLRS are stable over time, and multiple stable conductance levels can be programmed either by adjusting VG during the set process (which fixes RLRS) and/or by adjusting the end voltage of the negative RVS (which fixes RHRS, Extended Data Fig. 5).

5The most surprising observation, however, relates to the endurance, which readily reaches 2.5 million cycles (Fig. 2d–f) when applying sequences of pulsed voltage stresses. Under this type of stress, the values of RHRS, RLRS and RLRS/RHRS can be accurately controlled in three different ways: by tuning the duration of the write pulse, by tuning the amplitude of the write pulse, and by tuning the amplitude of the erase pulse (Fig. 2d–f and Extended Data Fig. 6). This endurance is very high considering the small size of the memristors (Supplementary Note 2), and similar to that of commercial metal-oxide-based resistive random access memories (0.5 million cycles) and phase-change memories (10 million cycles),. However, the switching time of the 1T1M cells using top Au–Ti electrodes is rather long (tSET of 232 µs and tRESET of 783 ns, Extended Data Fig. 7).

6The properties of the 1T1M cells can be adjusted using different top electrodes (Extended Data Fig. 8). When Au electrodes are used, the devices show reliable switching at lower state resistances (Fig. 2g), as well as shorter switching time (t) and lower switching energy (E), and when Ag electrodes are used, these values can be pushed down to tSET = 680 ns, tRESET = 60 ns, ESET = 21.11 pJ and ERESET = 1.41 pJ (Fig. 2h). The reasons behind these observations are the lack of an interfacial Ti layer (which is prone to absorb oxygen, increasing the out-of-plane resistance) and the higher conductivity and diffusivity of AuX+ and AgX+ ions (Supplementary Note 1). The performance observed in h-BN–CMOS 1T1M cells using Au–Ti electrodes may allow them to cover niche applications between NAND Flash and DRAM within the memory hierarchy (for example, persistent memory), and when using Au or Ag electrodes their performance may be valid for low-power application-specific integrated circuits within the internet-of-things (Supplementary Fig. 12).

7On the basis of the above measured performance metrics, the hybrid 2D–CMOS 1T1M cells show good potential for data computation. The high RHRS/RLRS ratio and the stability of the resistive states over time allows us to implement in-memory computing operations taking advantage of the internal connections of our 5 × 5 crossbar array of 2D–CMOS 1T1M cells. As a proof of concept, we realized ‘or’ and ‘implication’ operations (Extended Data Fig. 9), although more sophisticated operations could be easily realized by modifying the interconnections between the devices by means of custom design.

8Furthermore, the 1T1M cells with Au–Ti–h-BN–W memristors show spike-timing dependent plasticity (STDP) when applying pairs of pulsed voltage stresses displaced in time at the input and output (Fig. 2i). This non-volatile RS performance is very attractive to construct electronic synapses for spiking neural networks (SNNs), which consume less energy than traditional deep neural networks.

9Although implementing by means of hardware a reliable 2D materials-based memristive SNN capable of competing with state-of-the-art developments, is not yet achievable due to the lower maturity of these materials, we can analyse the performance of a SNN made of memristors that show STDP characteristics such as those in Fig. 2i (Supplementary Note 3). First, we fit the measured STDP data from Fig. 2i, including the device-to-device variability, using an exponential decaying model to implement the learning rule (Supplementary Fig. 13). Second, we simulate a SNN to demonstrate the unsupervised learning capability (Fig. 3a), and benchmark it by classifying the images from the Modified National Institute of Standards and Technology (MNIST) database of handwritten digits, (Methods). The SNN has 784 input neurons, an excitatory layer of 400 neurons and an inhibitory layer of 400 neurons, plus a decision block that determines which is the most probable digit (0–9) represented by the input pattern. We trained the SNN with the complete MNIST dataset and evaluated the accuracy every 1,000 images. Figure 3b–d shows the main three figures-of-merit for this type of SNNs (that is, evolution of the synaptic weights with the number of training images, confusion matrix of the network and the training accuracy versus number of training images) and all of them indicate an excellent performance. To account for the device variability, we considered a Monte Carlo simulation with 50 iterations that randomizes the exponential fitting of the STDP plot and the initial value of the synapses, and the deviations observed in the accuracy are very low (less than 5%, Fig. 3d and Supplementary Fig. 14). The best average accuracy reaches roughly 90%, which is a very high value considering the simplicity of the SNN and the unsupervised training protocol (Supplementary Table 4).

10We also propose a CMOS circuit for the hardware implementation of an electronic neuron based on our h-BN memristors (Fig. 3e), which is capable of accounting for the adaptative firing threshold and the refractory period after firing (pre- and postsynaptic traces and the evolution of the membrane potential, simulated by means of SPICE, in Fig. 3f,g).

11Very few commercial electronic products today already include 2D materials, and the ones that do (sensors, specialty cameras) use very low integration density (more than 100 µm2 per device)—because in larger devices the local defects in the 2D material are not so detrimental. Our hybrid 2D–CMOS microchips are still far from being ready for production, but we can safely claim that our work represents the highest performance and technology readiness level ever achieved in high-integration-density 2D materials-based electronic devices or circuits. The electrical characteristics of the h-BN memristors connected to a CMOS transistor are by orders of magnitude superior to those of standalone h-BN memristors,– and h-BN memristors connected to 2D materials-based transistors,.

12The voltages needed to switch our devices (from ±1.4 to ±5 V) are low compared to other prototypes in the field of 2D materials (even more than 20 V)–, but still higher than that used at the 180 nm CMOS node. Nevertheless, this is not an impediment for the development of this technology, as there are many commercial microchips that operate at much higher voltages; that is the case for all Flash memories (state-of-the-art 3D-NAND Flash memories are programmed at around 20 V) and all bipolar-CMOS microchips for automotive applications (which require up to 40 V). Strategies to fabricate wafers with devices that operate at different voltages are widespread, and many companies, offer versions of their 180 nm CMOS technology that operate at high voltages greater than 18 V. Note that prototype memristive devices developed by companies also operate at ±5 V (ref. ).

13We finally remark that, at a first glance, the use of Au and Ag electrodes may not appear ideal because they are categorized as contaminant in front-end-of-line (FEOL) processes. However, our h-BN memristors are integrated in the latest metallic layer of the BEOL interconnections (Fig. 1h), where Au pads, liners and wires are usually used (Extended Data Fig. 10). The semiconductor industry has also developed ferroelectric memories with high content of Iridium (a contaminant material forbidden in FEOL processes), and companies working in the field of 2D materials use Au electrodes in their studies and FEOL prototypes (Supplementary Table 5). Hence, the use of Au, Au–Ti or Ag electrodes in our hybrid 2D–CMOS microchips for memristive applications does not prevent their adoption by the industry.

14The metal-oxide-semiconductor field-effect-transistor circuits were fabricated in a standard CMOS foundry. The size of the wafers is 200 mm, and the technology node was 180 nm. Each wafer contained 60 chips with a size of 2 cm × 2 cm, and each of them contained different circuits, including the unfinished memristors, finished transistors and unfinished 5 × 5 crossbar arrays of 1T1M cells. As we received the silicon wafer without a passivation layer, first we cut it to separate the microchips and then etched the native oxide by immersing them in a diluted hydrofluoric acid solution (10:1) for 1 min to etch the native oxide (SiO2). This step was carried out to expose the conductive tungsten vias to make a good electrical contact with 2D materials. Second, we transferred a sheet of multilayer h-BN (previously grown by CVD on a Cu foil) using a wet transfer method. One layer of polymethyl methacrylate (PMMA) with thickness of around 300 nm was spin coated on the cut h-BN. The PMMA–h-BN–Cu sample was deposited on a FeCl3 solution (0.1 g ml−1) to etch the Cu substrate and, once the Cu disappeared, the resulting PMMA–h-BN sample was washed in diluted HCl solution (1 mol l−1 for 1 min) and deionized water (for 1 h). The PMMA–h-BN sample was picked up using the native oxide-free CMOS microchip and dried naturally in a dry box. The PMMA was then removed by immersing the sample in acetone for 24 h. Third, we used photolithography (mask aligner from SUSS MicroTec, model MJB4) to expose the h-BN on top of the metallic pad. Then we used a dry etching method with Ar–O2 plasma (Plasma Cleaner from PVA TePla America Inc., Model IoN 40) to etch the h-BN (300 W for 10 min) and expose the pads. Finally, we used photolithography, electron beam evaporation (Kurt J. Lesker, model PVD75) and a lift-off process (rinse in acetone for 1 min) to pattern and deposit the top electrodes and/or drain electrodes (3 nm Ti with 40 nm Au on top without breaking the vacuum, or 50 nm Au or 50 nm Ag). The process was simple and reproducible, although we believe it could be considerably improved if optimized methods in an industrial clean room are used. Ideally, the h-BN should be grown in large CVD systems (for example, Aixtron) and transferred on the wafers before cutting them in multiple microchips, using methods like laser debonding. Note that even large companies are still using small (7.6 cm or 3-inch) tube furnaces to grow the h-BN for their prototypes. We also confirm that getting the microchips finalized and etch the passivation film before transferring the h-BN works well.

15The morphology of the devices was investigated by an optical microscope (DM 4000M, Leica), AFM (Dimension Icon, Bruker) and TEM (Titan Themis, FEI). The thin lamellae for TEM inspection were prepared using a scanning electron microscope provided with focused ion beam (Helios G4 UX, Thermo Fisher Scientific). The electrical characterization was performed by using two probe stations (both M150, Cascade) connected to different semiconductor parameter analysers: a Keithley 4200 and a Keysight B1500A. All the IV curves under DC voltages were collected using the Keithley 4200 in the ramped voltage sweep mode, for which three source-measure units are needed for drain, source and gate. Also, all the IV curves under pulse mode were collected by Keysight B1500A with two Waveform Generator/Fast Measurement Units connected to the drain and source. An Agilent E3631A DC Power Supply was used to apply constant voltage stress on the transistor gate as gate voltage for pulse measurement. All the endurance plots were collected using the recommended method described in ref..

16The SNN architecture has been developed using Brian2 (ref. ), an SNN simulator written in Python. The learning process is based on the empirical measurement of STDP made in the 1T1M cells combining an Au–Ti–h-BN–W memristor on a CMOS transistor. We considered the variability of the network by running a Monte Carlo engine. We benchmarked the accuracy of the SNN during image classification of the MNIST dataset of handwritten digits under an unsupervised learning scheme. We propose a circuit-level model for the neuron–synapse–neuron system, as well as its implementation in SPICE. A detailed description of the SNN and its performance is given in Supplementary Note 3.

17Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-023-05973-1.