Neuromorphic Computing Primitives Using Polymer-Networked Nanoparticles Authors: Yinong Zhao, Xingfei Wei, Rigoberto Hernandez Published: 2024 ABSTRACT AbstractNanoparticle networks have potential applications in brain-like computing yet their ability to adopt different states remains unexplored. In this work, we reveal the dynamics of the attachment of polyelectrolytes onto gold nanoparticles (AuNPs), using a bottom-up two-bead-monomer dissipative particle dynamics (TBM-DPD) model to show the heterogeneity of polymer coverage. We found that the use of one polyelectrolyte homopolymer limits the complexity of the possible engineered nanoparticle networks (ENPNs) that can be built. In addressing this challenge, we first found the commensurability rules between the numbers of AuNPs and poly(allylamine hydrochloride)s (PAHs). This gives rise to a well-defined valency of a AuNP which is the maximum number of PAHs that it can accommodate. We further use an engineered block copolymer, which has a conductive middle block to mediate the distance between a dimer of AuNP. We argue that by controlling the length of conductive block that is connecting the AuNPs and their respective topology, we can have ENPNs potentially adopt multiple states necessary for primitive neuromorphic computing. Introduction Advances in conventional silicon-based computing materials, with increasing performance and decreasing feature sizes approaching that of indivisible atoms, have brought us to the turning point of their physical and practical limitations. As Moore’s Law reaches diminishing returns, we also approach the von-Neumann bottleneck that hampers data transfer efficiency,1−3 and the quest for alternative computing paradigms is becoming increasingly important. A promising approach to overcome these challenges lies in the development of neuromorphic computing encoded in neither silicon-based nor living materials. Thus, we seek novel materials that emulate the brain’s neural architecture to achieve its high computational and power efficiency.4−10 One such effort would be the development of nonsilicon-based memristors11,12 to store information in the form of electrical resistance. With such memristors, for example, the computation and information processing can then be achieved by connecting these individual units together into a giant network.13 Another example of the implementation of nonsilicon-based materials for information processing involves the use of nanoparticle networks14−18 with control of their optical and electrical properties through their shape and size. Earlier, we characterized nanoparticles with DNA linkers to form DNA origami19 that gives rise to unique structures of nanoparticle circuits.20,21 Meanwhile, others have demonstrated that percolating nanoparticle systems can be responsive to external stimuli.22−28 Among the candidates for such nanoparticle systems, engineered nanoparticle networks (ENPNs) stand out because they realize geometries that can mimic the synaptic connections in the brain.13,23,29−35 The connectivity and properties of such ENPNs through directed assembly of nanoparticles with various linkers can be precisely controlled at the molecular level,36−47 and thereby offer the possibility of the selectivity and precision necessary for creating neuromorphic systems. Here, we suggest the use of polyelectrolytes as a linker48 between the polymer and corresponding engineered nanoparticles (ENPs) through physical attractions and preselected commensurability. We define the valency of an ENP with respect to polymers of given lengths as introduced by Yao et al.,49 and also employed by some of us20,21 in describing the maximum accommodation of single stranded DNAs on an ENP. Specifically, for the combinations of a gold nanoparticle (AuNP) at a given diameter and many poly(allylamine hydrochloride)s (PAHs) of a given length (shown in Figure 1), the valency of the AuNP is the maximum number of those PAHs—as benchmarked in Section 3.1—that the AuNP admits on its surface. (a) Two-bead-monomer DPD (TBM-DPD) model representation with each monomer of PAH coarse-grained into a “N bead” (hydrophilic bead) and a “C bead” (hydrophobic bead). (b) In this model, both the intrinsic property of PAH and the heterogeneous coverage when interacting with AuNP are benchmarked by adjusting bond bending. (c) Heterogeneous coverage of PAH on AuNP is recovered for single and two chains cases, for both all-atom (AA) and TBM-DPD model. (d) Commensurability between different sized AuNPs and number of PAHs is demonstrated for three cases. Specifically, a 4 nm AuNP can accommodate two PAH200, a 6 nm AuNP can accommodate three PAH200, a 8 nm AuNP can accommodate six PAH200. The first hypothesis, verified below in Section 3.2, is that such a valency exists for preselected AuNP diameters and polymer lengths. Homopolymers that are sufficiently long fill the surface of an AuNP, and thus a sufficiently long polymer can be used to attach a pair of gold nanoparticles (AuNPs). This property also identifies the limitation of using PAH as a homopolymer to construct complex ENPNs. It can be circumvented if homopolymer segments can be isolated from the rest of the polymer, and if such binding segments can be used as partial binders to a given ENP. To address the latter, here we need a second hypothesis, verified in Section 3.3, that shorter polymers can fill the valency of a given ENP for a specified commensurate ratio of polymers to ENPs. To address the former, we need the third and final hypothesis verified in Section 3.4, that engineered triblock copolymers can utilize such commensurability between AuNPs and PAH for the attachment between dimeric ENPNs, and the middle block as a spacer eliciting controlled separation between the AuNPs. Taken together, ENPNs can potentially adopt preselected states as would be necessary in creating primitive materials for neuromorphic computing. Simulation Models and Methods The PAH and AuNP models are chosen because they exhibit heterogeneous structuring on their contacting interface. We coarse-grain the AA PAH model to develop a two-bead-monomer dissipative particle dynamics (TBM-DPD) model where each hydrophobic bead corresponds to the backbone repeating unit and each hydrophilic bead corresponds to a ammonium group attached to the backbone; see Figure 1a. We systematically tune the angle bending parameter and match the heterogeneity of surface coverage that are preserved from AA model to the coarse-grained dissipative particle dynamics (DPD) model; see Figure 1b,c. Using the TBM-DPD model, we investigate the spatial coverage of PAH on different AuNP surfaces. Based on such heterogeneity, different sized AuNPs are characterized depending on their ability to accommodate different numbers of PAH chains, leading to the valency indicated in Figure 1d that we demonstrate below. Coarse-Graining Model for DPD Simulations The coarse-grained models are mapped from all-atom simulation models, using the following steps: First, we obtain trajectories from all-atom simulations for AuNPs and PAH after equilibrium—either from our prior work if they are available or from our new simulations—where the force field is from Heinz and co-workers50 and OPLS,51 respectively. We then systematically mapped the all-atom models to the coarse-grained DPD representations. The same coarse-graining scheme was used earlier for modeling a single PAH wrapping onto a citrate capped AuNP,52 thus our method preserved all the relevant properties. In DPD AuNP models, they consist of 473, 1596, and 3784 coarse-grained beads for the corresponding AA AuNP models at 4, 6, and 8 nm diameter, respectively. We use the fully protonated PAH polymer linkers to establish a more universal benchmark for the system. Future studies will dive into the effects of varying protonation states on the PAH, to inform the parametrization of this polyelectrolyte under various pH conditions. The interactions between coarse-grained (CG) breads are be captured by the DPD force field53 The first term on the right is the electrostatic force, which is associated with the charges at i and j The remaining terms on the right are the conservative (neutral), dissipative, and random forces, respectively As in the usual implementations of DPD,53,54 the higher order terms are ignored, and FijC in eq 1 has been set to be a linearly repulsive force in eq 3 with a cutoff at Rc. We also assume that the conservative interactions can be classified with only two types of beads. The force constant aij in FijC takes on the value 25 between beads with similar hydrophobicity and 80 otherwise as is standard practice.55,56 The dissipative forces FijD in eq 4 are connected to the random forces FijR in eq 5 through the fluctuation–dissipation theorem at a given temperature, T. In the TBM-DPD model for PAH, the hydrophobic beads are charge-neutral, and the hydrophilic beads each carries one unit of positive charge. The latter have favorable electrostatic interactions with the negatively charged beads on the AuNP surface. In order to account for the heterogeneity in the surface charge density on the AuNP, the CG beads cannot all be of the same uniform charge. For simplicity, we take the limiting case of only two possible such beads: charge-neutral beads representing a subsurface of gold atoms functionalized to no more than a cutoff number of H2citrate–, and negatively charged beads representing a subsurface of gold atoms functionalized to more than a cutoff number of H2citrate–. In the case of the 4 nm diameter AuNP, at a typical surface ligand density, we find that the surface is mapped to 373 charge-neutral beads and 100 neutral beads. Similarly for 6 and 8 nm AuNPs, the number of negatively charged beads are 225 and 400, respectively. Using the above coarse-grained scheme DPD model, with the number density set to ρ = 3 and the volume of each bead set to 90 Å3,53 our DPD model yields a unit conversion of 6.46 Å to one DPD unit. In this model, the dielectric constant was set to ϵr = 78.3 to account for the aqueous environment. Since the system involves high charge density inherent to polyelectrolytes, one must dress the interactions in eq 1 when two beads come within their effective radii due to the soft potential prescribed by FijC in eq 3. Consequently, the electrostatic force FijE in eq 2 must be modified.57 The overall forces on the CG beads also include the bonded interactions between them, which we treat using a simple Gaussian chain model. Namely, the bond stretching and angle bending are simulated by harmonic potentials. All relevant simulation parameters are listed in Supporting Information (SI). Simulation Protocols All molecular dynamics (MD) simulations are carried out using the Large-scale Atomic Molecular Massively Parallel Simulator (LAMMPS) package.58 All the systems are run in explicit solvent conditions and periodic boundary conditions are enforced along all three box dimensions in a periodic box large enough that the solute part of each system is at least 20 DPD units away from their respective image. In practice, the box size was therefore 100 × 100 × 100 in DPD units. Consequently ca. 3,000,000 DPD beads—at the prescribed density—are required to fill the simulation box. All the polymers are initially fully extended and energy minimized before insertion into an empty simulation box. When polymers and AuNPs are inserted into the simulation box, they are spaced out by at least 2 nm in order to ensure that they are initially not in contact. Initial conformations for all simulations were specified and generated using Packmol59 and Moltemplate60 with salt ions added to neutralize the net charges in the system. Additional simulation details on system size and components are listed in the SI. Following the initialization, each trajectory is equilibrated to 300 K in the canonical ensemble and then run under microcanonical conditions for data production. In all simulations, AuNPs are treated as rigid bodies while the PAHs on the surfaces are allowed to move freely to reflect the nature of physical bonding and their dynamic attachments. Considering the system sizes and the need for sufficient data collection to ensure statistical relevance, a total of 5 trajectories—that is, replicates—were integrated for each configuration. Results and Discussion Benchmarking of the TBM-DPD Model We first benchmarked our TBM-DPD model to recover the relationship between chain rigidity and chain length. The rigidity is tuned by changing the angle bending harmonic spring constants on the backbone, with . We found that using kangle = 0.5 we achieved the radius of gyration (Rg) from the CG PAHs in good agreement with the AA PAHs at different chain lengths; see Figure 2.The behavior of Rg is often reported in the form of a power law where R0 is the prefactor, N is the degree of polymerization, and α is the exponent. From the least-squares-fit of eq 6 to the Rg in the TBM-DPD model, the exponent of the scaling relationship is obtained as α = 0.75. This is in moderate agreement with the reported value—viz α = 0.81—observed in experiment by Frank et al.,64 and appears to be in closer agreement than the value of α = 0.60 found computationally earlier63 which is to our knowledge the best previously reported such value. In the AA model, the PAH chain thus appears to behave as a polymer that moves in a good solvent. Interestingly, Wang et al.61 recently reported a related single-bead-monomer DPD (SBM-DPD) model having an α = 0.96, which, when compared with experiments, performed better at high degree of polymerization, but worse at low degree of polymerization. When considering the discrepancy between their model and experiment,61 it is worth noting that the Flory–Huggins theory employed in their DPD model to construct the force field parameters is a mean-field approximation that could oversimplify the interactions in the polymer model. For example, the Flory–Huggins approach assumes a homogeneous mixture and does not fully consider the detailed local interactions or specific chemical nature of the polymer or solvent. In addition, the Flory–Huggins theory typically assumes spherically uniform particles, which is therefore unable to distinct hydrophobic and hydrophilic regions. These differences can be captured with the more detailed TBM-DPD model in the present work because it includes a separation of charge between the “N” and “C” CG beads, as indicated in Figure 1a. Comparison of Rg versus the degree of polymerization obtained from simulations using either the DPD model presented here (red stars), or that of Wang et al.61 (blue diamonds) to several power-law forms—viz eq 6 with the optimized constants, R0 and α—as follows. The red and blue dashed curves are the least-squares fits to the corresponding simulations, with the goodness of fit characterized by χ2 and R2 having the values of 1.28, 0.986 and 0.03, 0.999, respectively. The other two power-law curves are generated using the optimized constants reported by Olenick et al.62 to the AA model data63 (yellow) and the experimental data64 (green). Surface Coverage of PAH as Homopolymer The surface coverage of a PAH homopolymer was first characterized for cases where it was connected to a single 4 nm diameter AuNP. The structure was initialized so that the polymer chain was extended (and nearly linear), and a central axis was defined by its two ends. A AuNP was placed along this axis in proximity to one of the polymer ends. In this single AuNP case, the average tail-to-COM distance for several PAHs of varying lengths as a function of simulation propagation time is reported in Figure 3a. Initially, the tail-to-COM distance reflects the chain length of the PAH in its relaxed and coiled state. After approximately 400,000 time steps, for PAH at all chain lengths except at PAH800, the distance from the tail of the PAH to the center of mass (COM) of the AuNP remains nearly constant at ca. 3.3 nm. This distance corresponds to the radius of the AuNP at ∼2 nm plus the CG bead size at ∼1 nm, suggesting that the monomers—that is, the DPD beads on the PAH—are all laying on the surface of the AuNP. This further indicates that for shorter polymer lengths (PAH200, PAH400, PAH600), the entire PAH chain binds to the AuNP surface. However, for PAH800, the tail-to-COM distance at equilibrium is significantly higher than the outer edge of the nanoparticle surface, indicating that some monomers of PAH begin to extend into the solution after completely covering the AuNP surface; see Figure 3b–d. The existence of a point at which the AuNP surface is fully coated—viz filled—by PAHs without any dangling polymer strand suggests that it has a well-defined valency for the commensurate ratio of PAHs to AuNP at 1:1. If we were to consider homopolymers longer than this commensurate length, then the remainder of the polymer would extend beyond the coated AuNP surface. (a) End-to-COM distance of PAH of different length wrapping on a 4 nm AuNP, results are collected and plotted from 5 distinct trajectories, with the shaded area indicating the error bar. The status of the AuNP monomer is categorized as “under-filled”, “filled”, or “over-filled” depending on the length of the respective PAH, as shown in (b–d). (e) COM-to-COM distance of PAH of different length bridging two 4 nm AuNP together, results are collected and plotted from 5 distinct trajectories, with the shaded area indicating the error bar. The status of the AuNP dimer are categorized into “collapsed dimer” and “extended dimer” depending on the length of the respective PAH, as shown in (f, g). For one homopolymer to fully fill two AuNPs, the polymer chain length would thus need to be longer than double the commensurate length. The remainder—or excess polymer subchain—would then be available to bridge the two AuNPs. For such dimers bridged with homopolymers, the COM-to-COM distances are reported in Figure 3e at different PAH lengths. When the excess polymer subchain is relatively short, the PAH beads collapse near the adjacent surface of the two AuNPs, bringing the two AuNPs into contact as shown in Figure 3f. When the PAH length increases, the two AuNPs begin to separate from each other with the middle part of the PAH chain extending into the solution; see Figure 3g. The PAH behavior in both single AuNP and dimer AuNPs cases show that the homopolymer would first fully fill the AuNP surface, then the remaining parts extend into the solution. Polyelectrolyte Valency Revealed on a Single AuNP Given the length of a homopolymer at a commensurate ratio of 1:1 to fill one valency, we now show shorter homopolymers, whose lengths are reduced by a factor of n, can combine together to fully fill many valencies at a commensurate ratio of n:1. For example, we previously observed a commensurate ratio of 2:1 in filling the valency of a AuNP using all-atom simulations.52 Therein, the first PAH200 was shown to cover a hemisphere of the 4 nm AuNP surface, leaving enough space for the second PAH chain to attach; see Figure 1c. Using the CG TBM-DPD model developed in this work, we observed such behavior in simulations with multiple PAH chains attaching to one AuNP at different sizes; see Figure 1d. We revealed the commensurability between AuNPs and PAHs, by parametrically varying the size of AuNPs, from 4 nm to 6 nm to 8 nm, with the PAH chain length held constant at PAH200. Following the protocol provided in SI for wrapping multiple polymers, we found that each PAH200 occupies roughly the same amount of footprint/surface coverage on the AuNP, regardless of the order in which they attach to the AuNP. Specifically, for 4, 6, and 8 nm different sized AuNPs, we found 2, 3, and 6 polymer chains can fully wrap onto the surface, respectively; see the left column in Figure 4. When attempting to attach the third, fourth, seventh chain on the 4, 6, and 8 nm AuNPs, respectively, we observed PAH either partially wrapping or completely detaching from the AuNP, see the right column in Figure 4. This behavior is due to the lack of empty space on the AuNP surface, as well as the electrostatic repulsion from the previously attached PAHs. Notably, the above result can be further confirmed from the percentage coverage of each chain. For example, each PAH200 occupies approximately 45% of the 4 nm AuNP surface, which sums up to ca. 90% of the surface with two chains attached, leaving insufficient room for a third one. Such trends can be observed in the case of 6 and 8 nm AuNPs as well. The commensurability then reflects the number of polymers—in this case, PAH200—that can be accommodated by a single AuNP given its size or diameter. Thus, for a 4 nm AuNP, the valency is found to be 2; for a 6 nm AuNP, the valency is found to be 3; and for an 8 nm AuNP, the valency is found to be 6. Surface coverage of PAH chains on a AuNP of different sizes. Top, middle, bottom rows are AuNP of 4, 6, and 8 nm diameter, respectively. Left column with panel (a, c, e) represents the maximum number of PAH a AuNP can accommodate according to their respective size, with the last PAH chain wrapping onto the surface in previous chains’ presence. Right column with panel (b, d, f) represents the failed attempt of another chain trying to wrap on a filled surface, when exceeding the maximum number of PAH accommodation. Using Block Copolymer to Control AuNP–AuNP Distance in a Dimer Previous simulations have demonstrated that the model lacks of control of the spacing between two bridged AuNPs, when using electrolytes—e.g., PAH—as a homopolymer. In order to overcome this challenge, we now consider triblock copolymers with PAHs at the two end blocks of the triblock copolymer as sticky ends and a middle block that is nonsticky. Doing so has the benefits of allowing us to control the length of the sticky blocks which is related to the degree to which they can bind to AuNPs and the middle nonsticky block as the spacer. In the current work, we demonstrate that using a triblock copolymer with two PAH sticky ends at length that gives a valency of 1 can construct dimers. Unlike in the case of homopolymers, using triblock copolymers as the bridging block has the benefit of the middle nonsticky block being nonattractive to the AuNPs. This ensures that there is no collapse of the middle bridging block, and the only question is whether the tribock copolymer spontaneously bridges between the AuNPs and what is the resulting distance between them. For specificity, we used a charge-neutral conductive polymer to be the middle bridging block because, in contrast to PAHs, it is not attractive to the citrate-coated AuNP surface. As we found above, the commensurate length of PAH540 gives a 1:1 valency for 4 nm AuNPs. Using two PAHs540 at the ends and varying the chain length of the middle block, the distance between the AuNPs monotonically increases with the chain length; see Figure 5a. Moreover, Figure 5b shows that the relationship between COM-to-COM distance Rcc of the AuNPs and the end-to-end distance Ree of the middle block of the copolymer both follow a power law similar to eq 6 with exponent approximately equal to 0.5; see Section S-4 in SI for details of the fits. This suggests that the middle block of the copolymer bridging the two AuNPs effectively behaves similarly to a Gaussian chain. In turn, the relationship between the AuNP-to-AuNP distance and polymer chain length is critical in gaining control of the spacing between AuNPs. For the present, this is a phenomenological (or top-down) result because we did not fully characterize the thermodynamics of this system, and its proof awaits future work. (a) COM-to-COM distances of AuNPs among dimers with different lengths of middle blocks. (b) Power-law fit of both the COM-to-COM distances, Rcc, between AuNPs and the end-to-end distance, Ree, of the middle block in the copolymer. R0 is the prefactor, N is the degree of polymerization of the middle block and Δ is the effective length of the two pseudo chains representing the distance between the COM and the point on the AuNP surface were the middle chain is tethered. The goodness of fit characterized by χ2 and R2 have the values of 0.28, 0.91 and 0.07, 0.99 for Rcc and Ree, respectively. Conclusions In conclusion, we developed a TBM-DPD model that is able to capture the characteristic interactions between PAHs and AuNPs. Using this model, we demonstrated the existence of a well-defined valency afforded by the commensurability between AuNPs and PAHs. Specifically, we found the commensurate length of the PAH at a 1:1 ratio for AuNPs at a certain size and the behavior of such homopolymers when bridging two AuNPs together. Then we use shorter PAHs to prove that multiple PAHs can be utilized as partial binders, which give rise to a multivalency of given sized AuNPs. Moreover, the use of a triblock copolymer is introduced with two PAH blocks at the ends wrapping onto AuNP surface and the conductive middle block mediating the AuNP-to-AuNP distance. In future work, we expect to take advantage of controlling the segment lengths of the middle block and the two end blocks on the triblock copolymer to design AuNP valencies with specificity and to construct emergent ENPNs. Such ENPNs were seen earlier to correspond to identifiable states that can potentially be used to encode information.65 The answer not resolved earlier, and now suggested here, is that we can use the valency of AuNPs to selectively construct the segments of ENPNs. An additional benefit of this choice is that the resulting ENPN can also be conductive when using conductive polymers, which is amenable to the use of electric conductivity in identifying—viz reading—the different ENPN states. As a simplest example, a triangular trimer of 4 nm AuNP can be readily designed with three triblock copolymers containing PAH200 outer blocks. Acknowledgments This work was supported by the National Science Foundation (NSF) through Grant No. CHE-2102455. The computing resources necessary for this work were performed in part on Expanse at the San Diego Supercomputing Center through allocation CTS090079 provided by Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS), which is supported by National Science Foundation (NSF) grants #2138259, #2138286, #2138307, #2137603, and #2138296. Additional computing resources were provided by the Advanced Research Computing at Hopkins (ARCH) high-performance computing (HPC) facilities. We would also like to thank Dr. Alexander Popov for insightful discussions. Data Availability Statement The data that supports the findings of this study are available from the corresponding author upon reasonable request. Supporting Information Available The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jpcc.4c06055. Detailed information about simulation model force field in Section S-1; a more detailed view of the surface coverage on the nanoparticles summarized in Figure 4 in Section S-2; and the forms and procedures for the fits in Figures 2 and 5 in Sections S-3 and S-4, respectively (PDF) Detailed information about simulation model force field in Section S-1; a more detailed view of the surface coverage on the nanoparticles summarized in Figure 4 in Section S-2; and the forms and procedures for the fits in Figures 2 and 5 in Sections S-3 and S-4, respectively (PDF) The authors declare no competing financial interest. Supplementary Material jp4c06055_si_001.pdf References Associated Data Supplementary Materials jp4c06055_si_001.pdf Data Availability Statement The data that supports the findings of this study are available from the corresponding author upon reasonable request.