Synaptic potentiation and depression in Al:HfO2-based memristor
Graphical abstract
Introduction
Despite the ability of solving well-structured problems, Von Neumann paradigm is inadequate to solve problems where inputs are not well defined, as in the real world [1], or to implement adaptive learning algorithms [2]. In this respect, one of the best solution is a neuromorphic kind of computation [2], [3], featuring high parallelism [3], [4], [5] and ability to perform adaptive learning [2].
The reference model of neuromorphic architectures is the living beings’ neural network, where neurons communicate by means of electrical or chemical signals passing through synapses [1], [5]. When synapses are stimulated by neuronal spikes, they present a behaviour called plasticity, i.e. they change their strength, thus either facilitating or inhibiting the connection between two neurons through potentiation and depression, respectively.
A device used as an optimum artificial synapse must therefore be able to continuously adapt its conductance from a Low Conductance State (LCS) to a High Conductance State (HCS) or vice versa upon application of electrical pulses.
Among the solutions to emulate biological synapses, an increasing interest is being devoted to Resistive Switching (RS) devices [1], [6], or memristors, which have been investigated over the last years for non-volatile memory applications [7], [8], because of their high scalability and low energy consumption [1], [2]. These two-terminal devices feature a change in their conductance upon application of proper electrical stimuli [1], [7] in the directions of conductance increase (SET process) and conductance decrease (RESET process). The same mechanisms can be exploited for neuromorphic applications to emulate, respectively, the potentiation (increasing conductance) and depression (decreasing conductance) in an artificial synapse [1], [6]. To this aim, several programming algorithms have been proposed, mainly based on the application of voltage pulses with different amplitudes [4], [9] or widths [5], [9] or even a train of identical pulses [3], [9], [10]. Among them, the latter algorithm is the most suitable for neuromorphic applications due to its ease of implementation [3], [9], [10] in a chip.
It has been reported that, while the RESET operation is usually gradual, the SET operation usually features an abrupt change in the device conductance, forcing to the use of an external current compliance [1], [2], [3], [10], [11], [12]. This additional external control is an obstacle for the implementation of the simple pulse scheme described above. Indeed, RESET operation only has been employed to emulate gradual synaptic depression [3], [10]. Furthermore, devices requiring current limitation cannot be considered analogue (even if multilevel operation has been demonstrated [10]). Only few works have demonstrated self-compliant [13], [14], [15] and analogue behaviour [5], [9], [14], [15].
In this context, we propose an Al-doped HfO2 memristive device operating in an analogue self-compliant way over a maximum conductance window of 10 on which different switching dynamics are investigated in pulsed regime. HfO2 is a material that guarantees compatibility with the processes of semiconductor industries and good performances when employed in RS devices [3], [4], [8], [10], [11]. Moreover, the doping of HfO2 with Al prevents the oxide crystallization, thus preserving an amorphous state, and improves device performances [2], [11]. The main reported switching for oxide-based RS devices is the filamentary one [16], [17], [18], consisting in the initial formation of a conductive filament inside the oxide (forming operation) and its subsequent partial disruption (RESET operation) and restoration (SET operation).
Section snippets
Materials and methods
The device area is 40 × 40 μm2, with a 40 nm TiN/5 nm Al:HfO2/10 nm Ti/TiN stack (Fig. 1, inset). Ti and TiN layers were deposited by magnetron sputtering and the Al:HfO2 (4% Al) layer was deposited by atomic layer deposition at 300 °C, as described elsewhere [19], [20].
The electrical tests were performed using the source measuring units (B1511B and B1517A) and pulse generator units (B1525A) of a B1500A Semiconductor Device Parameter Analyser by Keysight. The train pulse algorithm is a
Electrical testing
Initially, the device has a conductance of tens of nS. The conductive filament has to be created for the first time through a forming operation at around 2.5 V (data not shown), leading to a HCS (Fig. 1). To switch the device from HCS to LCS, and vice versa, DC sweeps from 0 V to −1 V (HCS to LCS) and from 0 V to 0.7 V (LCS to HCS) were applied (Fig. 1). The maximum conductance ratio that can be obtained is about one order of magnitude. The SET operation is therefore performed without the need of
Conclusion
We demonstrated that the fabricated Al:HfO2 memristor device can be employed as an artificial synapse, emulating the potentiation and depression processes by an easily implementable algorithm based on a train of identical pulses. Potentiation is proven to be obtained without current compliance. A careful choice of the combination amplitude–width is fundamental to emulate both potentiation and depression dynamics and it is more stringent for the potentiation than the depression. The memristor
Acknowledgments
This work was partially supported by the European project RAMP (FP7-ICT-2013-10). The authors acknowledge Dr. E. Cianci for the support in material synthesis.
References (26)
- et al.
Microelectron. Eng.
(2013) - et al.
IEEE Electron Device Lett.
(2011) - et al.
Nanotechnology
(2013) - Y. Wu, S. Yu, H.-S.P. Wong, Y.-S. Chen, H.-Y. Lee, S.-M. Wang, P.-Y. Gu, F. Chen, M.-J. Tsai, in: Mem. Workshop IMW...
- et al.
Adv. Mater.
(2013) - et al.
Sci. Rep.
(2014) - et al.
Nano Lett.
(2010) - et al.
IEEE Circuits Syst. Mag.
(2013) - et al.
Adv. Mater.
(2009) - C. Walczyk, T. Schroeder, M. Lukosius, M. Fraschke, A. Fox, D. Wolansky, B. Tillack, C. Wenger, in: Non-Volatile Mem....
Nanoscale
J. Electron. Mater.
Cited by (55)
Neuromorphic applications using MO<inf>x</inf>-based memristors
2022, Metal Oxides for Non-volatile Memory: Materials, Technology and ApplicationsPlasticity and learning behavior emulated in a ZnO based transparent artificial synapse
2021, Thin Solid FilmsCitation Excerpt :STP and LTP are essential characteristics of human memory; only a few types of artificial synaptic memories show STP and LTP in a single device. Almost all resistive random access memory (RRAM) shows synaptic behavior with a stable low resistance state (LRS) [18–20], and hence forgetting is absent. Forgetting is also a part of mammalian memory, which is very rarely observed in RRAM.
Short-term and long-term synaptic plasticity in Ag/HfO<inf>2</inf>/SiO<inf>2</inf>/Si stack by controlling conducting filament strength
2021, Applied Surface ScienceCitation Excerpt :The array operation was demonstrated in [20]. Long-term potentiation (LTP) and long-term depression (LTD) are most important features among many synaptic behaviors to implement hardware neuromorphic system [21-25]. Also, deep neural network (DNN) based on the backpropagation learning rule can be used for neuromorphic system.
Bipolar, complementary resistive switching and synaptic properties of sputtering deposited ZnSnO-based devices for electronic synapses
2021, Journal of Alloys and CompoundsCitation Excerpt :Regarding this, literature reports that multilevel characteristics are generally attained by adapting the LRS resistance when the CC is dictated through set process and HRS resistance via stop-voltage during reset-process [60–62]. It is also described in the literature that the irregularity of the resistances in the HRS and LRS, which is related to a conductive filament, depends on the various CCs [14]. Fig. 4(g) shows the possible switching mechanism of Ta/ZnSnO/TiN devices that could be described in terms of a gradual rise in filament size on increasing CC during various set operations.
Synaptic learning functionalities of inverse biomemristive device based on trypsin for artificial intelligence application
2021, Journal of Materials Research and Technology