NVM Selectors: A Unified Explanation of Threshold Switching

Photo of Ron Neale, Renowned Phase-Change Memory ExpertContributor Ron Neale joins us again to review a paper delivered at last December’s IEDM conference by John Moores & Cambridge Universities, IMEC, and the University of Wuhan.  While the main focus of the paper is on PCM endurance improvements, it also provides some new inputs, which, with some suggested additions of Neale’s own, might now provide a unified explanation of threshold switching in the chalcogenides.  Neale includes discussions of these new ideas with one of the paper’s authors.

One of the most interesting papers at the recent IEDM was presented by a team at: John Moores University Liverpool, and Cambridge University, UK, IMEC, Belgium and the University of Wuhan, China.  As its title makes clear, this research has an important target of improvements in the lifetime and endurance of thin film selectors.

Endurance improvement of more than five orders in GexSe1-x OTS selectors by using a novel refreshing program scheme
F. Hatem, Z. Chai, W. Zhang, et al.

However, the material which supports their success appears to offer from a single homopolar (Ge-Ge) bond something that might be much more important and exciting, which is a means of accounting for all of the switching characteristics of enriched amorphous selenium based selectors and even other compositions.

It is possible that this could offer a means of unifying the actions of: Threshold switching, Forming, Annealing, Switching Delay Time and some of the lifetime degradation processes, with the possibility of even more, if our readers will bear with us and allow with a little speculative effort of our own to extend the scope of the published work.  More of that later.

The starting point is the selector material which, in this case, is Ge-enriched amorphous selenium (Se), with the composition GexSe(1-x), where x = 40%. The selector device structure has a stack consisting of TiN- the amorphous GeSe -TiN as illustrated in the figure below, with a GexSe(1-x) film thickness of from 5nm to 20nm. The structure is configured with an extended planar upper electrode and a bottom electrode 50nm in diameter utilizing a 300nm fabrication process. A thin film of TiN is in contact with the active material separated from the bulk of top electrode by a stop layer.

TEM photo (left) and sketch (right) of a single cell from the IEDM paper

The details of the composition of the stop layer are proprietary, although the dark colour would suggest it is a good conductor (possibly carbon).  This also indicates that, while TiN is apparently needed as a contact material, any possible ingress of TiN into the GexSe(1-x) is minimized.

In its glassy state it is claimed GexSe(1-x) has within its structure certain useful defects, the form of which can be modulated by an electric field to allow the material to switch from an insulating to a conducting state. The defects of particular interest are formed by what are described as mis-coordinated homopolar (Ge-Ge) bonds.

When stimulated by the effects of an applied field, those bonds are able to effect a change in the Ge co-ordination number, resulting in more bonds associated with each Ge atom, until it becomes over co-ordinated.  In more simple terms the local bonding structure is modified so that the Ge atom has more than four atoms sharing bonding electrons with it.  All this made possible because of the flexible glassy matrix in which the Ge-Ge bond exists where reversible short-range structural changes can be easily accommodated.

More formally, supported in this IEDM 2019 paper by ab initio simulations and earlier work, the Ge-Ge bond in its initial state is considered to be localised or in a ground state.  The effect of the electric field results in what is called a de-localised or excited state.  In effect the de-localised states are larger, meaning they become closer together which, it is claimed, makes tunnelling between them easier.  This means the mobility increases which increases the electrical conductivity, i.e. the material is able to switch from being an insulator to a conductor.

When the electrical stimulation is removed the local bonding again becomes localised and the material returns to its insulating state.  All that is required to return the material to the insulating state is for the current to be reduced to some minimum level.  At that time most of the de-localised states, but not all, make a rapid return to their localized state until there are too few to maintain any current flow at the applied voltage Vc.

While there are insufficient de-localised states in the right positions to maintain the continuous percolation path, those which hang about in the de-localized state after the device has returned to its high-resistance state (formally described as de-localised slow defects) still have a role to play.  They are considered to be responsible for what is commonly described as annealing, which is the slow recovery of the threshold switching voltage to its higher operating values.  It was reported for the Ge Se devices that a full recovery back to the Forming voltage (Vff) takes about 40 days, with recovery to (Vff/2) in 1 microsecond.

Selectors are often defined by a Forming voltage and an operating voltage.  What this new work now claims is that for GeSe that difference is only really a matter of time.  Annealing is observed in both selectors and phase change memory devices (PCM).  Because nowadays the latter are usually fabricated in the conducting or crystalline state, it is difficult to know the value of the Forming voltage.

For selectors and phase change memory devices the threshold switching value is not fixed, nor is its variation with annealing time, as discussed in the earlier paragraphs of this article.  It also varies with over-voltage above some minimum value.  For each value of applied voltage there is a characteristic delay time before threshold switching occurs.  The delay time varies inversely with the applied voltage.  I would venture to suggest that the delay time may be due to the propagation of the short-range structural change of the Ge coordination number as the selector builds the initial percolation path from one electrode to the other.

The Missing Link.

Having accounted for Threshold switching, Forming, Annealing and Switching Delay time with a direct link to localized and delocalized states of the Ge-Ge homopolar bond and changes in Ge co-ordination number, is it possible to go one step further?  I think it is and here’s how: by using the de-localization of the Ge-Ge bond to account for the constant voltage Vc in the conducting state to provide the final piece in the threshold switching jigsaw. Many consider the constant value Vc is the signature of a constant current density filament as the current is increased and decreased.

Top: Current vs. voltage plots for device. Bottom, How conducting filament width increases to maintain constant current with varying voltage.

The observed characteristics of the threshold switch, illustrated in part (a) of the above figure, can be considered to be made up from two parts: a resistance component (yellow) which is electrically in series with the insulating and conducting states of the switch (blue).  The combined effects of these two parts give us the observed I-V characteristics (red).  Explanations for the constant voltage (Vc) suggest that it is the signature of a constant current density filament which expands and contracts in response to increases and decreases in current, as is illustrated by the two cross sections at the bottom of the figure.

The authors of the IEDM 2019 paper are clear that the conducting state is filamentary.  As the material undergoes threshold switching, the de-localised states form a percolation path to provide the conducting link between the two electrodes in the selector device structure (b) in the figure.

For a homogeneous GexSe(1-x) it film it must be assumed that the density of suitable Ge-Ge bond sites is constant, carrying with it the implication that the number of suitable percolation paths per unit area is also constant.

Therefore, once the initial conducting filament has been established at Vc, figure (b), any attempt to increase the voltage and current will find that there are more sites and new easy-to-form percolation paths in the material surrounding the conducting filament than within it.  These new sites can be more easily de-localized at Vc which results in the radial expansion of the filament at a constant Voltage (Vc) and the constant current density filament characteristics (c).  The final piece in the threshold switching jigsaw from one Ge-Ge bond site is in place.

From the published data my calculations of the current density indicate a range of from (5.1x 10^6) to (2 x 10^7) Amps/sq-cm for a conducting filament covering the whole of the bottom electrode.  If the filament is smaller in diameter, say 20nm, this rises to the range (3.15 x 10^ 7) to (1.26 x 10^8) Amps/sq-cm.  Irrespective of other effects it does suggest a significant increase in temperature, which means consideration must be given to the possibility that temperature could play a role in the de-localization process.  The easy answer to the question as to why Vc is a constant is because it is associated with a constant current density filament associated with any current.  However there may be other answers.

I raised this question with one of the IEDM paper’s authors, Prof. Weidong Zhang at the Liverpool John Moores University, inquiring about the diameter of the conducting filament.  His emailed reply was as follows:

“It is unlikely that the diameter of the filament can cover the whole area, because the on-current is the same in larger or smaller devices. This indicates that the filament diameter should be smaller than the BE area even in the smallest device.”

I then asked for his expert view on the radial expansion model of my figure, Prof. Zhang provided the following:

“We have no evidence at present to prove whether the diameter keeps constant at different on-current levels. An alternative explanation (to the filament expansion) is that defects in the filament at on-state have very low energy barriers, which can conduct very large current and do not need to expand the filamentary diameter.  So I am not sure that your diagram of filament expansion provides the one and only possible explanation.”

If the fixed-diameter filament model is correct, my concerns would be that any increases in current would move the current density into values of greater than 10^8 Amps/sq-cm.  This leads to considerations of the complex subject of permissible power dissipation in tunnel junctions, where it occurs, and to what level is acceptable.  I leave that to others for the moment.

It is not a new idea that changes in co-ordination number could be used to account for threshold switching.  One example was presented at IEDM 2014 by N. Takaura, et al in the form of a super cell structure consisting of layers of GeTe and Sb2Te3 to form a PCM.  The researchers claimed that changes in resistance were the result of electric field driven short-range motion of germanium atoms in the layers of the super lattice, utilising a crystal-crystal transition (i.e. second-order phase transition).  In that example the change in co-ordination number of Germanium was from six-fold for the low resistance state, to four-fold for the high resistance state.  The study used a structure considerably more complex than the more simple deposition of an amorphous film of Ge doped selenium.

Searching for the common denominator which links threshold switching in the millions of chalcogenide compositions where it is observed, I asked Prof. Weidong Zhang if the Ge-Ge homopolar bond was essential.  He replied:

“I don’t think Ge-Ge is a prerequisite for threshold switching. Other materials, so long they can produce similar structures, should also have the ability of threshold switching. For example, As/Te can also contribute to the traps in the mobility gap, which are supposed to be located at high-energy levels at off-state, triggered upon electron injection by high-bias pulses, annealed/disappear in time, just like the Ge-Ge bonds or over and under coordination pairs, i.e valence alternation pairs (VAPs).”

The subject of the IEDM 2019 paper was endurance, and the Ge-Ge bond has a role to play in the reduction of threshold voltage and increases in leakage current with switching lifetime. While there is no element separation with the single first switching event, careful TEM analysis provided conclusive evidence that Ge and (Ge-Ge) defects were moving towards the one electrode and Se towards the other with the accumulation of the number of threshold switching events.

The IEDM 2019 paper reported a series of experiments where sequences of 1k, 100k and 1M threshold switching pulses were applied before the application of a single reverse pulse.  After 1k switching events a full recovery of the device threshold voltage and leakage current was obtained.  For the 100k experiment a single reverse pulse could effect a partial recovery, while the 1M experiment recovery was impossible with just a single pulse.

For the 1M experiments the material was found to be both partially crystallized and porous.  This leads to the conclusion that while the device is in a condition where any degradation is recoverable, the situation is the result of more closely packed (Ge-Ge) homopolar bond structures resulting in a lower threshold switching values.  This is similar to what happens with leakage current, which increases as the number of switching events increases.  The ability to also recover from the degradation by thermal means partly supports the theory that the (Ge-Ge) homopolar bond structures were responsible.

In terms of endurance what the team discovered was that optimally, if a reverse polarity pulse was applied every 5k switching cycles, the effects of any degradation in threshold voltage and leakage current could be reversed.  This was accompanied by a claim that a single reverse pulse and thermal cycling could extend the switching lifetime up to 10^10 cycles.

The practical implementation of this in a memory array was not reviewed.  If the GeSe selector is used in an oxide memory cell (e.g HfO) the reverse pulses will be a normal part of the write/erase operation, although there will be some asymmetry in number of switching events for reading (more) than write/erase.  For a PCM cell it should be possible to implement a reverse-polarity write/erase, where the length the longer-duration lower-current SET pulse is of a different polarity to the short-duration higher-current RESET pulse.  Others have found that reverse polarity pulses are useful means of extending the W/E endurance of PCMs from element separation caused by electrostatic and electro-migration element separation effects. [See my review of IBM’s work on endurance.] This means reverse-polarity operation would aid in improving the life of both the selector and memory device.


In summary it would appear that it is now possible that, given the right defect structure to account for most of the characteristics associated with threshold switching (for example the Ge-Ge homopolar bond or something similar in a glassy amorphous matrix), the very viscous nature of the selector’s material allows the required reversible short-range atomic movement.

In the past it has been claimed that some types of homopolar bonds are the source of structural instability in, for example, GSS and GSSN compositions, and that these need to be annealed out [Reviewed in an earlier post in The Memory Guy].  Is it that very same structural instability which actually facilitates threshold switching characteristics?  Could it be that threshold switching is really just a matter of tinkering with the structure?

It would appear in any composition that the suitable defect structures will vary in type from those characterised as “de-localised slow defects” to those that are in the ideal position with respect to the applied field and more easily de-localised.  As far as instability of the material and the homopolar bonds, perhaps what is needed is a classification of the good, the bad, and the ugly.  Annealing removes the ugly along with those that only require a minimum of energy to disturb any stable situation.

That is not the end of the story.  Experimental evidence is needed of the way in which Vc, the conducting state constant voltage, varies with device thickness.  Is it a true constant?  There have been some observations that Vc varies as a function of composition suggesting a contact-like effect.

Another useful piece of evidence would the degree to which the constant voltage in the conducting state extends to where the material becomes molten, in particular for PCM.  If in the molten state the glass-like structural links or rings of Se and Te can be maintained then it is possible that the change in co-ordination numbers which account for the conducting state could also still occur.  If so then that might help resolve the “molten versus electronic switching” argument that in the past has plagued threshold switching.

As always of necessity I have perhaps oversimplified some key points for which I apologize.  However perhaps there are now enough pieces of the jigsaw available for those in academia to finally formalise a unified description of all aspects of threshold switching in amorphous chalcogenides.

One thought on “NVM Selectors: A Unified Explanation of Threshold Switching”

Comments are closed.