Carbon nanotube field-effect transistors (CNFETs) and their novel electronic properties have been the focus of intense research in the past few years 12. A key feature of these devices is the presence of large hysteresis in their transfer characteristics (I_DSV_GS curves) between forward and reverse gate sweeps. The hysteresis highly depends on the experimental and the environment parameters such as the gate bias range, gate sweeping rate, and temperature 34. Despite the difficulty to control it, hysteresis can be conveniently exploited to build simple memory devices 5678910111213.

In this paper, we study CNFETs where the single-walled carbon nanotube (SWCNT) has highly transparent contacts to the source and drain electrodes. We show that such transistors can work as memory devices with durability under continuous operation matching or outperforming endurance of Si-devices available on the market. The devices studied exhibit improved durability and charge retention at low temperature and low pressure. Measurements at room temperature show that such robust operation can be achieved also under ambient condition, without any passivation layer to protect the device from exposure to environment.

Figure 1 shows the SEM image of a typical device, consisting of an individual carbon nanotube embedded in Pd/Nb (2.5 nm/50 nm) contacts. Carbon nanotubes were grown on the substrate by a catalytic chemical vapour deposition technique, yielding clean and highly crystalline SWCNTs or DWCNTs 14. The leads were fabricated by photolithography, followed by metal sputtering and lift-off 15, and were separated by a ~2 μm gap. We used a thin Pd layer in direct contact with the nanotube to achieve high transparency contacts between the nanotube and the source and drain electrodes 16. The degenerately doped Si substrate (0.001–0.005 Ω cm) capped by 400-nm-thick thermally grown SiO₂ layer was used as back-gate.

Figure 2a shows the transfer characteristics of a device, displaying an ambipolar behaviour as a consequence of the small channel bandgap. At room temperature (blue triangle), the CNFET exhibits a small ON/OFF current ratio (≤10) and a low ON-state resistance (~20 kΩ) saturating at large negative voltage, confirming the high transparency of the contacts. The low ON/OFF ratio is due to the small SWCNT bandgap and can be enhanced by selecting a semiconducting SWCNT with smaller diameter and large bandgap. In our previous work 1517, we obtained devices with low contact resistance and large ON/OFF ratio, with the same fabrication process used for the samples discussed here. Because we use nanotubes with a small bandgap, we can test memory device properties for the same sample, with very different values of ON/OFF ratio, i.e. smaller ON/OFF ratio at room temperature or larger ON/OFF ratio at low temperature. Indeed, at low temperature (red circles), the ON/OFF ratio increases considerably: the current in the OFF state decreases and the current in the ON state increases. The increase in the ON-state current is due to reduced phonon scattering 1819. The ON-state current of ~10⁻⁶ A for a drain bias as low as 20 mV shows that the CNFET has a higher current drive capability than the present sub-micron Si-based devices 1620.

At a given V_GS, the hysteresis provides two distinct current values (ON/OFF states) that can be used as the logic levels of an easy-to-build memory device. The FET can be set to the ON state by a positive pulse at the gate (writing operation) and switched to the OFF state by a negative pulse (erasing operation). The states can be safely monitored at V_GS = 0 V (reading operation). An example of an erase-read-write-read sequence is shown in Fig. 2b.

The shortest width of the ±20 V single pulses that we used for memory switching is 5 ms. Although this switching time is two or three orders of magnitude larger than that of commercial flash memory devices 21, it is similar to switching times of CNFETs with SiO₂ gate oxide reported by other groups 622. We note that the switching time depends on the characteristic times of the charge traps in the gate oxide. It has been shown that CNTFET memory devices can reach switching times as small as 100 ns, using a different gate oxide 22. Here, we choose to focus on different properties that are important for data storage.

A memory device must have the robustness to withstand several write-read-erase cycles (endurance) and the ability to retain a certain state over time (retention). We studied these properties in our CNFETs at different temperatures and pressures. Figure 2c shows the data of the cycling test at 77 K and 10⁻⁵ Torr. The device exhibits remarkable endurance and maintains almost a constant programming window after undergoing ~1.2 × 10⁴ cycles. The transfer characteristic of this device does not show noticeable change after the test.

The retention times of both the ON and OFF states are tested by continuously recording the drain current in the ON (OFF) state at different temperatures and pressures. The results in Fig. 3d show that after a rapid increase in the OFF-state current during the first hour, the ON and OFF states remain well separated and the retention time exceeds 8–10 h in ambient air and under vacuum at both room temperature and 77 K. At low temperature, the current increase is slower and the separation of the ON/OFF currents is larger.

As described in 13, at least two mechanisms with very different time constants take place in neutralizing the positive charge stored at the SiO₂/SWCNT interface. The I_DSt curves can be explained considering both the shallow and the deep traps at or near the SiO₂/SWCNT interface. Charges stored in the shallow traps have a higher probability of escape by thermal field emission, while charges stored in the deep traps require larger gate electric field to de-trap. Hence, the deep traps effectively help maintain the long-term separation of the two current levels. In addition to the contribution from shallow traps, the rapid change that takes place during the first hour after write (or erase) pulse can be partially due to charges attracted from the surrounding air at the SiO₂/SWCNT interface, which neutralize the stored charge. For the latter reason, low pressure favours current stability as can be seen by comparing the I_DSt curves at 760 Torr and 3 × 10⁻⁴ Torr 1323. This suggests that for better device performance exposure to air has to be avoided. In addition, low temperature further improves retention by increasing the ON/OFF current separation.

We also tested the endurance of our devices in air at room temperature, as shown in Fig. 3. The data in Fig. 3a are for the device that was previously tested at 77 K (see Fig. 2c). While this device undergoes ~1.2 × 10⁴ cycles at 77 K and 10⁻⁵ Torr by maintaining a constant programming window (Fig. 2c), a steady degradation occurred after 10³ cycles in a successive test in air as shown in Fig. 3a. The inset in Fig. 3a shows that the window closure is triggered by the continuous reduction in the hysteresis width. This indicates a deterioration of the charge trapping mechanisms causing hysteresis after the stress test in air, probably due to the formation of additional defects on the oxide surface or to loss of adhesion of the SWCNT to the SiO₂. The extra defects may affect the current propagation in the nanotube and cause additional scattering, thereby decreasing the ON current at large negative gate voltage (see inset in Fig. 3a). We note that if the reading was done at a different gate voltage (for example 8–10 V), or if we monitored the threshold voltages rather than the current, we may have observed an almost constant programming window even beyond 8 × 10⁴ cycles.

The hysteresis degradation after the stress test in air does not occur in all the samples. We cycled a second SWCNT memory from the same production batch and with comparable channel resistance, kept at room temperature and atmospheric pressure. Figure 3b shows that the device has no obvious degradation after more than ~1.2 × 10⁵ erase-read-write-read cycles, outperforming the specification of the memory devices presently on the market. Therefore, we rule out the hypothesis that the change of hysteresis after the cycling test in air for the device in Fig. 3a is due to the ionization of air by the high electric field, as suggested in 5.

In summary, to our knowledge, the best endurance under continuous operation ever reported for SWCNT memory is 1.8 × 10⁴ cycles 22 before device failure. Our devices showed remarkable endurance beyond 10⁵ cycles in air, which can be considered as a breakthrough for CNT memory devices. We do not believe that the low ON/OFF ratio of our CNFET favours its endurance, since it increases the total amount of charge passed through the nanotube during the long cycling. Instead, the low contact resistance and the good quality of the SWCNT channel could prevent excessive self-heating of the SWCNT and its electric breakdown due to burning by oxidation (which usually happen above 600°C), as observed in other similar studies 2224.