An artificial synapse based on molecular junctions

The work targets the fundamental challenge of creating artificial synapses at the molecular length scale to improve integration density and energy efficiency beyond conventional CMOS and von Neumann architectures. Molecular electronics has recently advanced due to improved wiring methods and transport understanding, enabling molecular-scale switches, rectifiers, transistors, memories, and logic gates. However, most molecular devices show discrete, nonvolatile switching suitable for RRAM rather than the continuous, analog weight modulation needed to emulate synaptic plasticity. Neuromorphic computing requires hardware synapses with electrically tunable conductance to implement compute-in-memory paradigms. Prior memristive effects in molecular junctions have relied on intrinsic molecular states or molecule–electrode coupling, including redox and conformational changes, but continuous, reversible modulation for synaptic functions has remained elusive. This study proposes and demonstrates a molecular synapse whose conductance can be gradually and reversibly modulated via electrically driven silver cation injection into a peptide self-assembled monolayer, enabling both short- and long-term synaptic plasticity and basic temporal signal processing.

The authors situate the work within molecular electronics and neuromorphic hardware. Molecular junctions have exhibited electric-field-induced conductance switching based on intrinsic molecular characteristics (e.g., redox states, conformational changes) and electrode–molecule coupling. Such switching has enabled digital memory elements (molecular RRAM) but does not provide the analog, continuous weight changes required for synaptic emulation. Neuromorphic computing requires artificial synapses with dynamic, stimulus-dependent conductance changes to realize short-term plasticity (STP) and long-term plasticity (LTP). While memristors at larger scales are widely used to emulate synapses, implementing similar behavior at a monolayer molecular scale remains a key open problem. This work leverages an active Ag/AgOx electrode as a cation reservoir and peptide SAMs with coordinating groups to facilitate Ag+ migration and interaction, bridging gaps in prior literature by enabling analog, reversible modulation within a molecular junction.

Device architecture: A self-assembled monolayer (SAM) of an 11-amino-acid peptide (sequence CAAAAKAAAAK; C, A, K denote cysteine, alanine, lysine) is sandwiched between a bottom active Ag/AgOx electrode and a top eutectic gallium–indium (EGaIn) liquid electrode with a native GaOx skin, forming Ag/AgOx//peptide//GaOx/EGaIn. The cysteine thiol anchors to Ag; the top EGaIn contact is a conformal, soft contact without chemisorption. Fabrication: Bottom Ag films (200 nm) were deposited on Si by e-beam evaporation, then partially oxidized by annealing at 150 °C in air for 40 min to form Ag/AgOx and reduce the barrier for Ag+ injection. Template stripping with epoxy and glass slides yielded ultra-flat Ag/AgOx surfaces. Peptide SAMs were assembled from 0.5 mM ethanolic solutions for 12 h, rinsed, and dried. Controls included SAMs of dodecanethiol and other peptides (e.g., HS-PEG8-CH2CH2COOH; and C(GABA)(GABA)D(GABA)(GABA)D) assembled in corresponding solvents. For control devices without oxide activation, Ag electrodes (no anneal) were used. The GaOx/EGaIn top electrodes were patterned using PDMS microfluidic channels (five through-holes per PDMS slab) to contact the SAMs; approximate junction area ~ (10 µm)^2. Multiple junctions were fabricated by relocating the PDMS slab. Characterization: AFM in tapping mode assessed morphology and local roughness distributions; step-height AFM measured SAM thickness using masked regions. Angle-resolved XPS (AR-XPS) characterized Ag/AgOx and SAM chemistry; SAM thickness and surface coverage were extracted using established inelastic mean free path models and angular attenuation. Ellipsometry independently measured SAM thickness using optical models (Brendel for Ag, Tauc-Lorentz for peptide). Electrical measurements used a Keysight B2912A; pulses were applied to the top EGaIn via a Pt wire, with bottom Ag/AgOx grounded. Cyclic voltammetry (CHI660E) probed electrochemical behavior. Signal processing for reservoir computing used a DAQ (NI PCIe-6361) and SR570 preamplifier. Electrical protocols: Dynamic conductance modulation was probed by sequences of triangular and square voltage pulses. Example protocols: consecutive triangular pulses (amplitude 0.45 V, width 9 s) with positive-first then negative polarity; potentiation by negative pulses and depression by positive pulses. Potentiation sequences of square pulses (e.g., −1 V) were followed by monitoring current relaxation to extract double-exponential decay constants (τ1, τ2). Paired-pulse facilitation and spike-timing-dependent plasticity protocols varied pulse amplitude and inter-spike intervals to extract Δwt. Modeling: First-principles calculations examined electronic structure and Ag+ binding. Bare peptide electronic levels were computed (GGA-PBE) showing ~5 eV HOMO–LUMO gap; peptide–Ag+ systems were treated with wB97xd functional and Lanl2dz basis, revealing Ag+-induced in-gap states. Charge transport rates were estimated via Marcus hopping theory using computed hopping integrals and reorganization energies; Arrhenius-like temperature dependences of inverse hopping rates were compared to experiments. Continuum drift–diffusion simulations used coupled Poisson–Nernst–Planck (PNP) equations to model ionic (Ag+) and electronic transport under pulsed biases, reproducing potentiation/depression behaviors and double-exponential relaxations (weak chemical gating by free Ag+ vs stronger coordination binding). Reservoir computing: Devices were operated with alternating write (potentiation) pulses (−1.0 V, 1 s) and read (depression) pulses (−0.25 V, 10 s) over 120 cycles to assess stability. A simple RC scheme used device nonlinearity and fading memory to classify temporal inputs (sine vs square waveforms). Performance metrics included normalized root mean square error (NRMSE) and classification accuracy as a function of mask length; training and testing sets each contained 250 waveforms.

• Molecular-scale synapse: A peptide SAM junction (Ag/AgOx//CAAAAKAAAAK//GaOx/EGaIn) exhibits continuous, reversible conductance modulation under electrical stimulation, enabling synaptic weight tuning at monolayer length scale.
• Polarity-dependent modulation: Positive triangular pulses (0.45 V, 9 s) produced little change; subsequent negative pulses of the same magnitude caused a dramatic current increase (~20× after 75 pulses). Reversing polarity restored low-conductance states, enabling repeated potentiation/depression cycles (endurance demonstrated over ≥65 cycles with 30-pulse segments).
• Mechanism: Under negative bias, Ag/AgOx releases Ag+ that injects into the peptide SAM. Ag+ interacts with peptide amide groups (−C=O⋯−NH), enabling long-distance cation migration. Two contributions govern dynamics: (i) chemical gating by free Ag+ (fast relaxation, τ1), and (ii) coordination binding to peptide sites (slow relaxation, τ2). Control devices without AgOx activation or with long alkyl SAMs lacked Ag+ activation/migration and showed no similar modulation.
• Double-exponential relaxation: After potentiation with 20, 60, 100 pulses (−1 V), current decays fit double exponentials I(t)=a·exp(−t/τ1)+b·exp(−t/τ2)+c. τ1 increased with the number of potentiation pulses (more injected Ag+), while τ2 remained relatively constant, consistent with weak vs strong binding modes.
• Threshold and nonlinearity: A potentiation threshold near −0.3 V was observed; Δwt increased nonlinearly with pulse amplitude beyond this threshold. Below −0.3 V (e.g., −0.2 V), no appreciable conductance change.
• Short-term plasticity (STP): Paired-pulse facilitation (PPF) observed as positive Δwt=(I2−I1)/I1 for closely spaced pairs. Δwt decreased with increasing pulse amplitude and decayed exponentially with increasing inter-pulse interval (0.1–2 s), mirroring biological synapses.
• Long-term plasticity (LTP) via STDP: Spike-timing-dependent plasticity exhibited potentiation for Δt<0 and depression for Δt>0, with Δwt=(It−I0)/I0 measured 10 s before/after spikes (at 0.15 V). Exponential fits yielded time constants of ~350 ms (Δt<0) and ~529 ms (Δt>0), reflecting longer τ2-driven coordination effects.
• Modeling agreement: DFT showed Ag+-induced in-gap states reducing activation energies and enhancing hopping rates. Marcus theory reproduced temperature trends of inverse hopping rates. PNP simulations captured potentiation/depression curves, voltage dependence (−0.6 to −1.5 V), and double-exponential decays; without Ag+, no potentiation/depression occurred.
• Stability and statistics: Over 120 write/read cycles (−1.0 V, 1 s; −0.25 V, 10 s), devices showed reproducible behavior without degradation. Extracted τ1 and Δwt distributions were approximately Gaussian across cycles.
• Reservoir computing performance: Using device nonlinearity/fading memory, classification of sine vs square waveforms achieved 100% accuracy at a small mask length; NRMSE and accuracy trends vs mask length were quantified (ten devices).

The results demonstrate that a monolayer molecular junction can serve as an analog artificial synapse, addressing the need for continuous, reversible weight modulation at the ultimate scaling limit. The Ag/AgOx electrode functions as a cation reservoir, and the peptide SAM provides coordinating sites and pathways for Ag+ migration, enabling dynamic modulation of electron hopping. The interplay of fast chemical gating by mobile Ag+ and slower coordination binding explains both short-term and long-term plasticity, including PPF and STDP, with measured time constants consistent with distinct relaxation processes. Modeling from first principles and continuum PNP transport supports the proposed mechanism and reproduces key experimental behaviors. Integrating these molecular synapses into a reservoir computing framework validated their utility for temporal signal processing, achieving perfect classification of simple waveforms at small mask lengths. These findings indicate that biomolecular components can provide neuromorphic functionality with potential advantages in energy efficiency and integration density.

The study presents the first experimental realization of a molecular artificial synapse at one-molecule-length scale, using a peptide SAM between Ag/AgOx and GaOx/EGaIn electrodes. The device exhibits continuous, reversible conductance modulation driven by electrically injected Ag+ ions, enabling emulation of key synaptic functions (STP via PPF and LTP via STDP). The underlying mechanism combines rapid chemical gating and slower coordination interactions, as supported by DFT/Marcus analysis and PNP simulations. The synapse operates reliably over many cycles and serves effectively as a reservoir for simple waveform recognition with 100% accuracy at small mask lengths. Future directions include extending to single-molecule devices, exploring richer peptide chemistries and ion species, scaling arrays for more complex neuromorphic tasks, and interfacing biomolecular electronics with biological neural networks for sensing and modulation.