Sie sind auf Seite 1von 2

news & views

MEMRISTORS

Going active
The spiking phenomena associated with neural activity are characterized by an impressive degree of efficiency. The fabrication of a neuristor consisting of nanoscale components represents a step towards implementing such devices in integrated circuit applications.

Wei Lu

dvances in electronics have greatly impacted the human society over the past several decades. The main driving force behind these advances is the continued down-scaling of device size, resulting in higher speeds and larger packing densities at a lower cost for each new generation of device. To sustain this scaling trend, industry is constantly exploring new architectures and physical mechanisms in the search for the next device breakthrough. These advances also bring about the possibility of changes to the computing schemes themselves. Writing in NatureMaterials, Pickett and colleagues1 present such an example, by combining two-terminal devices known as memristors to build an active neuristor circuit that can perform several key neural functions. First coined by LeonChua in 19712, memristors can roughly be thought of as resistors with memory. Related devices commonly known as resistive random-access memory have already been pursued by the semiconductor industry for next-generation memory applications. Memristors differ from other two-terminal devices in that their states cannot be determined solely from the instantaneous inputs. Instead, they depend on how signals are applied over time and hence they exhibit memory effects. In 2008, Strukovetal. first linked the resistive switching phenomena in thin-film devices with Chuas memristor model3 and created a unified framework for device and circuit developments. Memristors are particularly well suited to neuromorphic systems because they can provide the plasticity and connectivity that allow the construction of biologically inspired computing circuits4,5. However, in previous studies, memristors act only as passive elements, in that they enable computing by modulating the connections between active units, but do not directly inject energy or amplify signals themselves. Pickett and colleagues demonstrate a functional, memristor-based active circuit capable of injecting energy and amplifying signals. More importantly, the circuit effectively emulates the action potential generation and propagation along an axon

Cell body

Axon

Figure 1 | Emulating an axon. Top: schematic diagram of a neuron. Bottom: spike generation and propagation along a chain of neuristors (N), mimicking the action potential propagating along an axon.

in three respects (Fig.1): an all-or-nothing threshold action, a refractory period during which the action potential cannot be repeated, and a constant propagation velocity and wave shape. These properties were first identified and proposed as central to the operation of a device HewittCrane termed the neuristor in 19606. Crane further argued that complete logic can be achieved with networks of neuristors, although like many concepts proposed ahead of their time, research on these devices never gained sufficient traction due to the rapid advances of transistor-based digital logic. To build a neuristor, three components are required for each unit: an energy source, an energy-storage element and a negative resistance device. Pickettetal. implement these components by using a battery, a capacitor and a memristor, respectively. As the memristor switches between the highresistance and low-resistance states activated by the input voltage pulse, the capacitor is sequentially charged by the battery and discharged through the memristor, in turn raising and lowering the voltage within the circuit. Consequently, a second memristor downstream can be switched and subsequent charging and discharging events are generated, producing voltage spikes at the

output when the two memristorcapacitor pairs are biased properly. Significantly, because the spikes are purely determined by the internal circuit parameters such as the charging and discharging time constants, bias voltages and coupling strength between the memristors, the output waves naturally exhibit a constant shape with a refractory period. Although prototype neuristors were demonstrated soon after their original proposal7,8, what sets the current study apart is that the memristors that make up this circuit are much more scalable, robust and energy efficient. Furthermore, resistance switching in memristors can now be extremely well characterized by means of numerical simulations1. Taken together, these properties make it feasible to build largescale memristor-based circuits. Strictly speaking, most two-terminal resistive switching devices, including the device used in the present study, do not follow Chuas exact definition of a linear memristor model2 and should be categorized as memristive systems instead. In particular, the device fabricated by Pickettetal. has a temporary memory, and the switching is better characterized as threshold, or volatile, switching instead of memory, or non-volatile,
93

NATURE MATERIALS | VOL 12 | FEBRUARY 2013 | www.nature.com/naturematerials

2013 Macmillan Publishers Limited. All rights reserved

news & views


switching. Nevertheless, the device has a well-defined internal state variable that can be treated within the formalism of memristive systems. To emphasize the temporary memory effect, the authors use the term Mott memristor. This work is clearly significant because a number of applications such as signal repetition, active transmission and other neuristor logic can be readily implemented as its result. Along with recent developments in memristor-based synapses, this study also opens new opportunities for memristorbased neuromorphic systems. On the other hand, it should also be kept in mind that a neuristor is not a neuron, and it can only implement a subset of neuronal functions. However, given the rapid developments in memristor research, it seems likely that more advanced circuits will soon be feasible9: recent studies have shown that memristors can not only emulate biological components phenomenologically, but they can also mimic various ionic dynamics realistically at similar energy and timescales5,10.
Wei Lu is at the Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, Michigan 48109-2121, USA. e-mail: wluee@eecs.umich.edu References
1. Pickett, M.D., Medeiros-Ribeiro, G. & Williams, R.S. Nature Mater. 12, 114117 (2013). 2. Chua, L.O. IEEE Trans. Circuit Theory 18, 507519 (1971). 3. Strukov, D.B., Snider, G.S., Stewart, D.R. & Williams, R.S. Nature 453, 8083 (2008). 4. Jo, S.H., Chang, T., Ebong, I., Bhadviya, B.B., Mazumder, P. & Lu, W. Nano Lett. 10, 12971301 (2010). 5. Ohno, T., Hasegawa, T., Tsuruoka, T., Terabe, K., Gimzewski, J.K. & Aono, M. Nature Mater. 10, 591595 (2011). 6. Crane, H.D. IRE Trans. Electronic Comp. EC-9, 370371 (1960). 7. Cote, A.J. Proc. IRE 49, 14301431 (1961). 8. Nagumo, J., Arimoto, S. & Yoshizawa, S. Proc. IRE 50, 20612070 (1962). 9. Chua, L., Sbitnev, V. & Kim, H. Int. J.Bifurcation Chaos 22, 1230011 (2012). 10. Chang, T., Jo, S.H. & Lu, W. ACS Nano 5, 76697676 (2011).

Published online: 16 December 2012

STRUCTURAL GLASSES

Flying to the bottom

It has been shown that glasses prepared by physical vapour deposition have extraordinary stability. A computer algorithm that mimics such a process has now identied the optimal deposition temperature and the glasses structural features.

Giorgio Parisi and Francesco Sciortino


lasses retain in their structure information about their history and the way they were initially prepared. To understand this, it is useful to think of glasses in the context of a potential energy landscape a complex corrugated multidimensional surface composed of basins of different depths and widths1,2 (Fig.1, left). In this energy landscape, a liquid at equilibrium is continuously exploring different regions, whereas a glass is trapped in a specific basin, the height of which is related to the glasss properties. The lowest basin for a disordered state would thus correspond to a glass that has the strongest cohesive energy and optimal thermal stability. When cooling a liquid to form a glass (that is, at conditions at which the nucleation of the thermodynamically stable crystal phase is bypassed), the faster the cooling rate is, the higher up in the energy landscape the glass is trapped and the more similar its structure is to that of a hightemperature liquid. So far, glasses cooled at the slowest experimentally accessible cooling rate have however remained trapped in basins significantly higher than those that are supposed to lie deepest in the energy landscape. Glasses have also been prepared by destabilization of the crystal structure, or by physical vapour deposition of atoms or molecules on a cold substrate. But as with the fast-cooled glasses, these methods also resulted in trapped configurations that
94

are high in the energy landscape. In fact, by progressively depositing particles on a substrate the generation of a disordered structure reminiscent of that of hightemperature liquids would be expected. This is why the recent discovery that exceptionally stable organic glasses can be produced by vapour deposition when the deposition rate is low and the deposition temperature is high (significantly higher than in previous studies) came as a big surprise3. Neutron refractivity experiments on these ultrastable glasses showed that molecules lying at the top of the deposited film had enhanced mobility 3, which is suggestive of the existence of a mobile (liquid) surface layer constantly regenerated by the incoming flux of molecules (once deposited, a layer of molecules becomes glassy in a few seconds). Apparently, in the same way the surfaces of crystals are not always so crystal-like, the surfaces of glasses may not be so glassy 4. Now, Singh, Ediger and dePablo report in NatureMaterials that ultrastable glasses can also be reproduced by a computer algorithm that mimics physical vapour deposition, and that these model glasses lie extremely low in the energy landscape5 (Fig.1). The algorithm, which is a modified molecular dynamics simulation, progressively introduces small groups of particles into the system while locally minimizing the potential energy and slowly reducing the temperature of the added particles. For years, scientists have struggled to devise algorithms capable of sampling the

regions of the energy landscape that lie close to the absolute minimum for disordered states. Singh and collaborators suggest that depositing particles from the vapour phase at the right conditions seems to be the most efficient way to sample the landscapes deepest basins. The authors also investigated the mechanism behind the formation of these model low-energy glasses at the singleparticle level, unambiguously proving that a liquid layer does indeed exist on the surface of the deposited glass, and showed that particles in this layer are characterized by a mobility that is several orders of magnitude larger than that in the bulk. Importantly, they also show that the packing in these glasses is remarkably uniform (with regular Voronoi polyhedra abounding as structural motifs). The statistical properties of the energy landscape close to its bottom, that is, the number and height distribution of the deepest basins, are intimately connected with the highly debated existence at low temperatures of a diverging correlation length in the glass. If no divergence exists, the free-energy barriers among the basins would always be finite and the potential energy would smoothly approach its low-temperature limit (Fig.1, right). Yet if a diverging correlation length exists, the barriers become infinite at a non-zero temperature (the so-called Kautzmann temperature, TK). At this temperature the glass would reach the deepest basins and the derivative of the potential energy would

NATURE MATERIALS | VOL 12 | FEBRUARY 2013 | www.nature.com/naturematerials

2013 Macmillan Publishers Limited. All rights reserved

Das könnte Ihnen auch gefallen