F. R.
Duarte
*a,
S.
Mukim
ab,
M. S.
Ferreira
ab and
C. G.
Rocha
cde
aSchool of Physics, Trinity College Dublin, Dublin 2, Ireland. E-mail: duarteff@tcd.ie
bCentre for Research on Adaptive Nanostructures and Nanodevices (CRANN) & Advanced Materials and Bioengineering Research (AMBER) Centre, Trinity College Dublin, Dublin 2, Ireland
cDepartment of Physics and Astronomy, University of Calgary, 2500 University Drive NW, Calgary, Alberta T2N 1N4, Canada
dHotchkiss Brain Institute, University of Calgary, 3330 Hospital Drive NW, Calgary, Alberta T2N 4N1, Canada
eInstitute for Quantum Science and Technology, University of Calgary, Calgary, Alberta, Canada
First published on 12th November 2024
Random nanowire networks (NWNs) are interconnects that enable the integration of nanoscopic building blocks (the nanowires) in a disorganized fashion, enabling the study of complex emergent phenomena in nanomaterials and built-in fault-tolerant processing functionalities; the latter can lead to advances in large-scale electronic devices that can be fabricated with no particular array/grid high-precision pattern. However, when various nanowires are assembled to form an intricate network, their individual features are somehow lost in the complex NWN frame, in line with the complexity hallmark “the whole differs from the sum of the parts”. Individual nanowire materials and geometrical features can only be inferred indirectly by attempting to extract information about their initial conditions from a response function measurement. In this work, we present a mathematical framework that enables inference of the intrinsic properties of highly complex/intricate systems such as random NWNs in which information about their individual parts cannot be easily accessed due to their network formation and dynamical conductance behaviour falling in the category of memristive systems. Our method, named misfit minimization, is rooted in nonlinear regression supervised learning approaches in which we find the optimum parameters that minimize a cost function defined as the square least error between conductance evolution curves taken for a target NWN system and multiple configurational NWN samples composing the training set. The optimized parameters are features referent to the target NWN system's initial conditions obtained in an inverse fashion: from the response output function, we extract information about the target system's initial conditions. Accessing the nanowire individual features in a NWN frame, as our methodology allows, enables us to predict the conduction mechanisms of the NWN subjected to a current input source; these can be via a “winner-takes-all” energy-efficient scheme using a single conduction pathway composed of multiple nanowires connected in series or via multiple parallel conduction pathways. Predicting the conduction mechanism of complex and dynamical systems such as memristive NWNs is critical for their use in next-generation memory and brain-inspired technologies since their memory capability relies on the creation of such pathways activated and consolidated by the input current signal.
For the case of the NWNs studied here, memristive junctions are organized in a network, they may compete with one another for activation during a linear ramp-up current evolution process; this can lead to a highly selective conducting state in which just a few junctions in the network take the load of propagating most of the current being sourced in the system. As in a mechanism to minimize power dissipation, under specific conditions, this state can evidence a single conducting path formed of nanowires arranged as memristive units in series. This selective conducting strategy aligns with artificial intelligence neuromorphic algorithms such as the “winner-takes-all” (WTA) implemented in computational neural network models, offering an in materia counterpart to the algorithm where the memory states are stored in the variable resistance states of the NWN.30,33–36 The use of adaptive path formation in NWNs for neuromorphic applications has been demonstrated experimentally and described theoretically; Loeffler et al.31 have shown via simulations an analogous to “synaptic metaplasticity” in the brain and the memory capabilities of NWNs when conductance path formation is reinforced in their skeleton; Diaz-Alvarez et al.37 combined experiments and simulations to describe the emergent dynamics and current pathway formation in NWNs using a robust atomic switching model that was able to elucidate prime neuromorphic features such as collective memory response, contrasting ON-OFF resistance states, and fault-tolerant capabilities. In subsequent work, Diaz-Alvarez et al.38 demonstrated an associative memory device out of NWNs by controlling their multiple resistance pathways. Remarkably, Zhu et al.32 took advantage of such NWN path formation dynamics to successfully demonstrate online learning from the NWN spatiotemporal dynamics by subjecting the networks to perform image classification and memory recollection tasks. In a noteworthy review article, Kuncic and Nakayama27 highlight various works that focus on the use of NWNs as neuromorphic hardware, with emphasis on their dynamical interconnectivity pattern of current path formation (including WTA states) and adaptability to time-dependent electrical stimuli relevant to applications in neuromorphic computing.
The WTA connectivity paths formed in NWNs are critical to defining the independently addressable memory or conductance states of the system. The formation of a WTA in a NWN system depends on various factors such as the characteristics of the junctions, the materials composing the core–shell nanowires, the overall network connectivity, and the junction density of the NWN. Because the phenomenon depends on so many correlated factors, conditions, and variables – some of them not directly accessible experimentally – it is key to conduct computational simulations to shed light on the main mechanisms that trigger (or not) WTA states. Manning et al.33 provided experimental confirmation and a modelling framework for WTA conducting pathways in NWNs formed with Ag nanowires coated with polyvinylpyrrolidone (PVP). To experimentally verify that a NWN would form a single WTA path, the authors used passive voltage contrast scanning electron microscopy imaging technique that depicted the wires being ‘short-circuited’ in the WTA path with distinguished colour contrast. This tedious imaging process, however, requires the activation of the system at least up until the formation of the WTA path. It would be useful to predict if a NWN will develop a WTA path or not prior to any excitation and in the absence of sophisticated microscopy imaging methods – as if one could foresee “the future” of the NWN. Moreover, Manning et al.33 performed extensive work on Ag/PVP NWNs which are prone to WTA states, but there is a wide range of core–shell material combinations that can be tested for WTA; a predictive computational method that could anticipate the emergence of WTA, before experimentation and for a wide variety of material specs, would enable the direct correlation of the initial conditions and intrinsic characteristics of the NWNs with its final (potential) WTA state. For that, one needs to rely on an inverse methodology that can decode the final WTA conductance state by inferring the inputs and initial conditions that lead to such a state. This inference is done with minimum information about how the NWNs are initially prepared. In this work, we present an inversion scheme capable of extracting NWN junction parameters that are crucially linked to the formation of WTA as well as parameters that are useful for the overall electrical characterization of the network junctions. We apply a computational inversion method, previously tested in disordered low-dimensional materials,39–42 that was able to estimate the concentration of scattering impurities based solely on the electronic conductance response of the system without any prior knowledge about its composition. The same principle is now adopted to uncover intrinsic properties of systems of higher complexity such as the NWNs. By accessing the NWN conductance response evolution, this method can extract information regarding the intrinsic properties of the system, e.g., the main parameters that rule their memristive junction properties, which, in turn, are associated with the emergence (or not) of WTA-conducting states. Consequently, we are able to provide a robust computational framework that can directly complement experimental microscopy means to characterize NWNs, including their material and overall network features.
To summarize our motivation, this work is guided by the challenges often encountered by computational descriptions when investigating real-world disordered nanomaterials that do not exhibit highly symmetric crystalline structures, impeding for instance, the use of periodicity arguments such as in Bloch's theorem. Amidst this challenge, we offer a computational framework that can infer the a priori unknown intrinsic properties of highly disordered nanomaterials. We introduce a novel supervised learning-based methodology that can extract unknown or inaccessible information out of disordered or amorphous types of materials, such as random NWNs. This will enable the inference or estimation of their intrinsic features theoretically. As demonstrated in Bellew et al.,43 the properties of a single interwire junction can be assessed in a simplified setup of only two nanowires in contact forming a single junction unit. This simplified configuration characterizes a very reduced “sample” out of a “population” of junctions when the nanowires form a network. The single junction experiment certainly provides an initial assessment of a junction, however, this may not characterize the complex environment when many junctions are interlinked in a NWN. Therefore, this research was driven to deliver a framework that can inform experiments conducted in disorder types of nanomaterials in which complexity and emergent effects can play a role in their collective responses. This framework, founded in an optimization scheme of a cost function (the misfit conductance function), can extract the intrinsic characteristics or initial conditions of the nanowires in a network environment and can predict how the NWN will evolve in time, following an inverse modelling fashion explained in detail in this work.
A pair of nodes in different wires that share the same contact defines a “junction” – the region of contact between two wires, and the point where the memristive properties of the dielectric coating are relevant. These junctions are described by memristive units whose resistance (Rj) changes as a function of the amount of current flowing through them and the history of the voltage applied across them. We can verify experimentally that certain nanomaterial systems whose transport properties are dominated by junctions display self-similarity, which is to observe that each nanowire junction possesses similar power scaling to that of the collective behaviour of the NWN junctions.33,34 Following the model used in our past works, the interwire junction resistance Rj or its reciprocal conductance Γj = 1/Rj can change from a low resistance state (LRS) to a high resistance state (HRS) via an activation process ruled by a phenomenological power law given by Γj = AjIαj in which I is the current flowing through the junction and Aj and αj are power law parameters that characterize the intrinsic properties of the junction. It is worth mentioning that our power law junction model has cut-offs in which the junction resistance cannot be larger than RHRS ∼ 104 kΩ and cannot be smaller than RLRS = 12.9 kΩ which is equivalent to the resistance quantum Rq = h/2e2 with h being the Planck constant and e the elementary charge. These cut-offs were established based on previous experimental results in which 104 kΩ is already a sufficiently high resistance for the kinds of dielectric coatings considered and 12.9 kΩ is based on the assumption that the amount of current compliance set in the experiments is enough to form one single conductance quantum channel inside the junction. The junction power law model with cut-offs as discussed above for fixed αj values and scanned as a function of Aj and I (source current) can be visualized in Fig. 3 as colourmap surfaces. The three panels correspond to the cases in which αj = 1.3 (top panel), αj = 1 (middle panel), and αj = 0.7 (bottom panel). In our previous works,33,34 it was observed that αj fluctuates around 1, therefore, one can view the case of αj = 1 as a critical point. This criticality aspect will be even more evident when we discuss the collective conductance obtained for NWNs (not only a single junction), and we will get back to it in subsequent sections. This figure shows clearly the regions of high and low conductance for a range of Aj at a fixed exponent αj. We can see that when αj < 1, the activation of the junction, i.e., when Γj starts to increase above the cut-off ΓHRS = 1/RHRS, initiates at lower current values but its conductance growth is slower or sublinear. Setting αj ≥ 1 will result in the junction activating at higher currents but its conductance growth will be faster, linear and supralinear, respectively. These features in the single junction model will render complex emergent outcomes when incorporated in a NWN frame and it is key to the realization (or not) of WTA states. In the next section, we will recap how the evolution of a NWN, formed by multiple power law conductance junctions, takes place when subjected to a ramp-up current source. This recap will serve to elucidate the necessity of an inverse modelling scheme. As we will see, the main issue is that the relationship of the single-junction parameters Aj and αj with WTA states in NWNs is not well-defined until the NWN has evolved to a known conductance power law due to the self-similarity observed between single junctions and NWNs.33,34 We will build a computational framework that allows us to predict the types of junctions that compose a NWN, meaning to predict Aj and αj from the NWN equivalent conductance and to anticipate if the NWN will develop a WTA state or not at the very early stages of its evolution, i.e., when the network is still activating and there is no sign of a power law being established.
![]() | ||
Fig. 4 Log–log plots of the NWN conductance evolution as a function of the sourced current; the NWN geometry adopted here is shown in Fig. 1. The junction power law conductance model parameters (given in arbitrary units) used in (a) are for fixed Aj = 0.10 and three distinct curves taken for αj = 0.8,1.0,1.2. In (b), Aj was reduced (and fixed) to Aj = 0.01 and the same αj exponent values as in (a) were used. In all figures, “a.u.” stands for ‘arbitrary units’. |
To computationally generate enough network samples so that we could perform the parametric simulations in this work, we extracted the typical wire lengths, nanowire densities, and network sizes from real NWNs synthesized in laboratory and devised a random NWN generation algorithm that would supply us with different synthetic NWNs while keeping essential similarities to the experimental ones. The main nanowire characteristics set in our model are detailed in the Results and Analysis section. A total of Nc = 100 synthetic samples was generated for each parameter characterization. The NWN conductance versus current evolution curves were obtained for all synthetic NWN samples from which the conductance averages taken over all CAs were calculated.
The conductance evolution curves are a result of the NWN memristive-based dynamics formulated in terms of the junction power law model introduced earlier. In this model, all junction conductances in the NWN are not static, rather they vary in accordance with the phenomenological power law model given by in which (Aj,αj) are parameters that depend on the intrinsic properties of the junction, and In,m is the amount of current flowing between the two terminal nodal points in the junction (n,m). The conductances of the nanowire segments are considered static given by Γin = 1/Rin with Rin being the nanowire inner resistance. We assume the nanowire metal cores are composed of the same material with a fixed inner resistivity of ρ = 22.6 nΩ m (relative to the silver nanowires) and the nanowire diameters fixed at 60 nm.
The NWN conductance evolution simulation starts by setting the initial conditions for all junction resistances at RHRS and considering that a minute amount of current (Imin) is being initially sourced in the metallic electrodes, electrically probing the NWN. Accounting for all NWN connectivity information and Kirchhoff's circuit laws, one can build the conductance matrix Γ and apply it to Ohm's law written in matrix form as
![]() | (1) |
![]() | (2) |
This junction current-flow value is then used to update all junction conductance values using the formula which will then be re-inserted in
Γ for the next evolution step. Note that Γn,m values are bounded by the reciprocal limits [RLRS,RHRS] to emulate the fact that junction conductances cannot increase or decrease indefinitely. The next evolution step is therefore conducted by incrementing the source current by a sufficiently small value of ΔI and repeating the process from equation (1). The NWN conductance evolution is carried on until a predefined maximum current source (Imax) value is reached. How large Imax is will determine if the NWN evolution will cover all transport characteristic regimes as explained earlier named OFF-state, transient growth (TG), power law (PL), and post-power-law (PPL) or just a subset of them. The CAs taken from the built training sets are obtained in such a way that all transport characteristic regimes are evidenced in the NWN conductance evolution.
Now, let us consider a partial conductance evolution curve up to the TG regime from which the NWN (Aj,αj) values are concealed or unknown. This truncated evolution constitutes the target signal of our inversion problem Γtg, and, by obtaining the junction model parameters (Aj,αj) we can infer if the NWN under study is prone to WTA conduction or not. For that, we will describe the χ-minimization scheme devised for the context of memristive NWNs ruled by power law junction models as illustrated above. We start by calculating the CAs which are the averages of the complete conductance evolutions. To infer the power law exponent value, we calculate CAs for the set of networks possessing a systematically varied αj value. In the misfit eqn (3), we calculate the discrepancy between the (truncated) target evolution Γtg of a NWN system and the CAs 〈Γ(I,αj)〉 and integrate within the evolution limits [Imin,Imax] as
![]() | (3) |
The computationally generated NWNs for the CAs resemble realistic NWN samples since their materials and structural features are based on experimental micrograph images from past works.3 While computationally generating the NWNs, the main spatial parameters and constraints set for the pseudo-random generator are:
• Length of electrodes and distance between them; those were set to create squared shape NWNs of 20 × 20 μm.
• The length of nanowires sorted in a given configurational sample is taken from a Gaussian distribution of average length 5.5 μm ± 2 μm standard deviation.
• The total number of nanowires was set to maintain the chosen standard of 900 wire–wire junctions (Nj), representing approximately the average junction density of the observed micrograph images. The number of placed wires also has to guarantee that the electrodes are electrically shortened.
• Each NWN configuration has a unique spatial distribution with nanowires being randomly placed at each generation.
All NWN samples computationally generated per the above recipe were evolved following the conductance evolution model – detailed in the Methodology section – for various αj and Aj values. For each set of (Aj,αj) parameters, the average conductance was computed from all NWN configurational groups for each current stage in the evolution, resulting in various 〈Γ〉 versus current curves. These curves for the whole training set generated are presented in Fig. 5(a) in which we use a colour band representation covering the range of all curves with a given αj value. The dashed and dotted lines on the same plot are the conductance versus current curves for two target NWN systems that we wish to determine their a priori unknown αj values. Note that the conductance evolutions depicted for the target systems only cover up to the TG stage and we do not know if these evolutions will result in αj values greater, equal, or less than 1. For this reason, we apply the misfit minimization scheme to unveil the unknown exponent αj adopted in each target simulation. Using the training set curves to compute the conductance averages plus the target conductance curves into eqn (3), we obtain the log(χ) as a function of αj curves appearing in Fig. 5(b). The current limits of integration used in eqn (3) are consistent with including only the OFF and the TG regimes in the χ cost function. Note that χ as a function of αj curves evidence clear minima. The dashed curve shows a minimum at αmin = 1.1, agreeing with the exponent of the target conductance curve that significantly overlaps with the yellow/green bands in Fig. 5(a). Similarly, the dash-dotted curve shows a minimum at αmin = 0.9, while overlapping with the saffron/orange bands. Another interpretation for the misfit minimization is that χ tracks the mean squared error between the target conductance and various average conductance curves created as the training set for distinct αj values. The ensemble in the training set that minimizes this error reveals the exponent of the target system.
![]() | ||
Fig. 5 (a) Colour band representation depicting the collection of average conductance versus current curves generated for the training set used in the misfit minimization scheme. Each collection refers to an ensemble of Nc = 100 synthetic NWN samples evolved using the junction power-law conductance model with cut-offs with Aj = 0.05 and distinct αj values ranging from 0.7 up to 1.6 represented by the depicted colour code. The colour bands populate the region between two consecutive ensemble average curves for given αj values. The dashed and dotted lines are target conductance evolution curves taken for different NWNs evolved up to the TG stage. We do not know a priori their αj exponent. (b) log(χ) versus αj curve computed using eqn (3) to determine each curve minimum. The dashed curve on panel (a) results in the black dashed log(χ) curve with its minimum at αmin = 1.1, while the dotted curve on panel (a) results in the dash-dotted red line showing a minimum at αmin = 0.9. In both cases, the correct exponent used in each target conductance curve was captured by the minimum value, even though that did not reach the power law regime in the evolution. |
We repeated the same χ calculation for other concealed αj = 1.1 NWNs as targets to make sure the αj prediction is not impacted by different networks of other spatial/configurational characteristics. For this, we considered a distribution of χ values computed from Ntg = 60 samples, where Ntg is the number of different target NWNs sharing the same αj used to define misfit function distributions (mean and standard deviation). In this way, we can analyze how χ fluctuates with respect to changes in the NWN spatial configuration as shown in Fig. 6. Once more, we can see the presence of a well-defined minimum with reasonable error bars depicting the χ fluctuations. This second identification of a minimum in χ characterizes an inversion calculation scheme; from the output response function of the system, in this case, the conductance as a function of current, one can predict an intrinsic parameter or initial condition of the system. Here, the exponent αj = 1.1, an intrinsic parameter in the conductance power law with cut-offs model, gives the minimum of χ, coinciding with the concealed target value. This is a supralinear exponent (αj > 1), as a result, one can expect the formation of a WTA even though the target function did not reach the power law stage yet. To verify this, one needs to explicitly visualize current pathways, i.e., to visualize how the NWN is distributing the source current through its network structure. This can be done through current colour mapping plots as in Fig. 7. In the figure, nanowire segments carrying substantial amounts of current are depicted in brighter colours. Each current colour mapping depicts a state of the NWN in the evolution curve of Γ versus I; for the results in Fig. 7, we fixed I = 0.25 (arbitrary units) which corresponds to a relatively early stage in the evolution as shown in Fig. 5(a). To demonstrate the contrast between a WTA and a non-WTA (multiple paths) situation, we obtained current maps for αj = 0.9, αj = 1, and αj = 1.1 for a NWN of fixed spatial structure as depicted in Fig. 7(a). In the cases of αj ≤ 1, we observe various nanowire segments being accentuated, indicating that the current sourced in the electrodes is being distributed to distinct sections of the NWN. For αj > 1 however, a single current pathway is highlighted characterizing the WTA state, which the misfit method could correctly predict without the need for current visualization maps.
![]() | ||
Fig. 7 (a) The same NWN as in Fig. 1 but only the nanowire segments that can carry current or are electrically activated are shown. This means that isolated nanowires or nanowire ‘dead ends’ are not depicted. (b)–(d) Current maps obtained at the conductance state at I = 0.25 which corresponds to a state at the transient growth (TG) regime (see Fig. 5(a)). The power law prefactor was fixed at Aj = 0.05. Each panel shows the current color map for a given αj value: (b) αj = 0.9, (c) αj = 1.0, and (d) αj = 1.1. In the panels, α ≡ αj. αj > 1 evidences a WTA scenario whereas in the cases where αj ≤ 1, a multiple path scenario is seen. |
Despite the fact that we emphasize the formation of WTA paths in random NWNs, it is important to note that depending on the characteristics of the junctions, WTAs may not develop as depicted in panels 7(b) and 7(c). WTAs are very interesting because they can be viewed as somehow counterintuitive considering the immense number of possible parallel paths a random NWN may have at its disposal to propagate current. Nonetheless, WTAs mostly form when the exponential characteristics of the junctions are supralinear, i.e., αj > 1. When the latter is not fulfilled, our simulations show a multi-path configuration as indicated in the current maps in 7(b) and 7(c). These states are also interesting to study since they are more representative of the emergent properties of the NWNs. In other words, while a WTA path is composed of a number of resistors in series, meaning that most of the equivalent resistance contribution comes from the summation with Ri being the resistance of the ith resistor in the WTA path and NWTA the number of resistors in the path, a multi-path configuration is not simply the sum of independent resistors, meaning that many junction resistances are evolving in a more collective/interactive environment. During the WTA evolution, some junctions evolve rather rapidly due to their supralinear exponent and local connectivity, whereas in the multi-path state, junctions evolve more collectively, as a result, our framework may be even more critical to characterize multi-path NWNs than WTAs in a complexity science point of view.
It is important to note that the misfit inversion procedure is not limited to infer only the exponent parameter αj. The conductance evolution of a NWN subjected to a current source depends on numerous other parameters such as its connectivity pattern (e.g., wire density that can define how sparse or dense a NWN is), junction density, material resistivity, nanowire length, etc. The procedure can be adapted to infer other variables as well, as we will show for the case of the power law prefactor Aj. Eqn (3) can be rewritten in terms of Aj as the parameter to be inferred from the minimization scheme, i.e., χ is a function of Aj instead. Another way is to conduct a two-dimensional scheme in which the minimization is done in terms of both parameters Aj and αj as we extended in.42 This bi-parametric inversion requires the creation of NWN CAs in a two-dimensional phase space of αj × Aj in which the computed χ points can be depicted as a surface contour plot. We sweep this space with Nc = 100 CAs having Aj values ranging from 0.03 to 0.09 and αj values ranging from 0.6 to 1.6 and applied the multivariable version of eqn (3) to calculate χ. The result is shown in Fig. 8, evidencing a log(χ) minimum at (αj = 1.1, Aj = 0.05), matching the power law model specifications of the target NWN system. Slicing this surface at αj = 1.1, we can visualize the minimum of χ as a function of Aj with clarity at Aj = 0.05 as shown in Fig. 9. In the latter, the χ values are averages taken over Ntg = 60 different target NWN inversions from which we also computed the standard deviations shown as error bars in the plot.
![]() | ||
Fig. 9 Fixing αj = 1.1 in Fig. 8, one can visualize the misfit function (average) plotted in terms of Aj; the minimum is located at Amin = 0.05, corresponding to the original prefactor set in the target NWN system. The average of χ was computed for Ntg = 60 different target NWNs with an accuracy of 95% and standard deviations shown as error bars. |
Having established the efficacy of the inversion method in characterizing the exponent and the power law prefactor modelling nanowire junctions in NWNs, we can construct a more realistic NWN case to test the misfit inversion scheme by considering that the NWN is made of two types of junctions modelled as: (1) Γ1j = A1jIα1j and (2) Γ2j = A2jIα2j. To focus this analysis on the exponents as a first approximation, we will keep A1j ≡ A2j and α1j ≠ α2j with α1j and α2j in the sublinear and supralinear regimes, respectively. This modelling consideration accounts for the fact that some junctions may activate faster (higher exponents) than others due to various microscopic factors, e.g., differences in the ionic mobilities of the charge carriers inside the junctions, differences in sizes of the nanowire coating layers, to name but a few. Fig. 10 shows the misfit function result to identify these two parameters from a target NWN with α1j ≡ α1 = 0.9 and α2j ≡ α2 = 1.2. Note the presence of the two equivalent minima – they arise because there is no differentiation between the indexes of α1 and α2 from the point of view of the calculation. In this case, both configurations of α1 = 0.9 and α2 = 1.2 and α1 = 1.2 and α2 = 0.9 are equiprobable.
The bimodal calculation above can be expanded to consider more than two exponents in the power law model, increasing the complexity of the NWNs, but as the phase space dimensionality of the problem increases, the computational cost for the misfit procedure (and the requirement for more datasets) also increases. Nonetheless, the misfit inversion procedure can be adapted to predict not all the individual αj's considered in the model, but the main parameters characterizing their probability distribution functions used to assign all αj's for all junctions. For instance, let's consider that the NWN target system is now built with a diverse set of junctions in which each αj value is randomly assigned following a Gaussian distribution of mean αμ and standard deviation ασ. These quantities are a priori unknown. We also construct NWN CAs in terms of the Gaussian distribution parameters that will be used in the minimization of the misfit function in eqn (3) written in terms of αμ and ασ. Fig. 11 shows a log(χ) surface as a function of αμ and ασ for a NWN ensemble of Nc = 300 and number of junctions Nj = 900. Once more, the dashed lines crossing marks the minimum in log(χ) found at αminμ = 1.2 and αminσ = 0.2. This last example demonstrates that the misfit minimization method can effectively predict parametric distribution features that characterize more complex NWNs with a wide variety of junction characteristics and material properties within the same NWN structure. In other words, this parametric distribution feature analysis allows the investigation of NWNs that are not as idealized as the first examples studied in this manuscript with all nanowire junctions with the same exponent and prefactor values, enabling hence the inference of more realistic settings as many NWN intrinsic properties and initial conditions can fluctuate significantly from sample to sample.
The prediction of αj, in particular, is very important in the context of memristive NWNs because that gives information on the conduction strategy the NWN is prone to use: it may use a single current pathway to propagate nearly 100% of the sourced current if αj > 1 or it may spread the sourced current into multiple paths if αj ≤ 1. The former is identified as the “winner-takes-all” (WTA) conduction state found in certain core–shell random NWNs studied in our group in the past.33 The misfit function methodology presented here is capable of predicting which exponent a target NWN system will evolve to by only using information from its early activation stages as if the NWN “future” can be anticipated from data acquired way before the NWN reaches the power law regime, being that the training set for the method. We also demonstrated that the methodology can be extended to make parametric predictions in two-dimensional phase space and can infer features of probability density functions that can model collective aspects of the NWNs in distribution form. As the dimensionality of the parametric phase space enhances, the computational cost associated with the methodology also increases, however, the misfit characterization of probability density functions offers a way to balance the computational cost by inferring collective features packed in probability density functions rather than all individual (distinct) variables characterizing, for instance, hundreds of interwire junctions. It is worth mentioning that the orders of magnitude of all parameters (floating or fixed) studied in this work are within the experimental range of NWN samples as studied in Manning et al.33 which were made of core–shell silver-PVP nanowires. We worked to fit exponents that are within the range of their findings, i.e., αj ≈ 1, prefactors Aj ∼ 0.1 (in arbitrary units), current sweeps in the nanoampere range, silver resistivity chosen from past experimental results at ρ ≈ 22 nΩ m, conductances in the range of the conductance quantum when they are activated (ON state) and in the hundreds of Ohms when deactivated (OFF state), and NWN sizes within experimental range of 20 × 20 μm. In this way, we managed to constrain the optimizations conducted in this work within experimental acceptable values.
This work demonstrates that the misfit minimization method can be applied to make key predictions in highly complex/disordered systems such as random NWNs exhibiting dynamical memristive behaviour, standing as a powerful mathematical tool for materials characterization when detailed microscopy experimental methods are not available or the initial conditions of the system under study, that responded to a certain input stimulus, are unknown. This strategy can be crucial when designing NWNs to meet certain adaptive conduction target functions that can serve future neuromorphic technologies.
This journal is © the Owner Societies 2024 |