🔬 applied physics

Stochastic Dynamics of Diffusive Memristor Blocks for Neuromorphic Computing

This paper investigates the synaptic convergence of three diffusive memristor-based spiking neurons through numerical modeling and experiments to identify voltage combinations for targeted activation and analyze spiking statistics, thereby advancing the development of universal functional blocks for neuromorphic computing.

Original authors: Wendy Otieno, Alex Gabbitas, Debi Pattnaik, Pavel Borisov, Sergey Savel'ev, Alexander G. Balanov

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wendy Otieno, Alex Gabbitas, Debi Pattnaik, Pavel Borisov, Sergey Savel'ev, Alexander G. Balanov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a tiny, brain-like computer. Instead of using silicon chips that think in strict "on" and "off" switches, this team of researchers decided to build a circuit that behaves more like a living brain: a little bit messy, a little bit noisy, and full of surprises.

Here is the story of their work, broken down into simple concepts.

The Characters: Three "Artificial Neurons"

In a real brain, neurons (nerve cells) talk to each other. Often, several neurons send signals to one single neuron, which then decides whether to fire a signal of its own. This is called synaptic convergence.

The researchers built a tiny electronic version of this using three "artificial neurons."

The Tools: They used special electronic components called diffusive memristors. Think of these as tiny, smart resistors that can change their resistance (how hard it is for electricity to flow) based on how much voltage you push through them.
The Magic: Unlike standard computer parts that stay the same, these memristors are "volatile." Inside them, tiny metal atoms (like silver) float around like a cloud. When you apply electricity, they drift together to form a temporary bridge (a filament) that lets current flow. When you stop the electricity, the bridge dissolves. This makes them act very much like biological synapses, which are also temporary and changeable.
The Setup: They connected two of these neurons (Neuron 1 and Neuron 2) to a third one (Neuron 3). Neuron 1 and 2 receive input voltages (let's call them $V_1$ and $V_2$ ). Neuron 3 listens to the "chatter" from the first two and decides whether to fire a spike (a burst of electricity) or stay quiet.

The Experiment: Tuning the Volume

The researchers treated the input voltages ( $V_1$ and $V_2$ ) like volume knobs on a stereo. They turned these knobs up and down, testing thousands of different combinations to see what would happen to the three neurons.

They asked: If I set the volume on the left and right speakers to these specific levels, which neurons will start "singing" (spiking)?

The Discovery: A Map of Behavior

They found that the behavior of the system wasn't random chaos; it was actually a structured map. Depending on the combination of voltages, the system fell into one of eight distinct "modes":

Silence: No one spikes.
Solo Acts: Only Neuron 1 spikes, or only Neuron 2, or only Neuron 3.
Duets: Neuron 1 and 2 spike together, or 1 and 3, or 2 and 3.
The Trio: All three neurons spike at once.

The "Noise" Factor:
Real brains are noisy. There is always a little bit of static. The researchers found that their electronic neurons were naturally noisy too, thanks to the random movement of the metal atoms inside the memristors.

Low Noise: The system was very picky. It only spiked if the voltages were just right.
High Noise: The system became more "generous." The random fluctuations helped the neurons fire even if the voltage was slightly too low. It's like how a little bit of background noise in a room might actually help you hear a whisper better by triggering your brain's sensitivity.

What Can This Do? (The "Brain Math")

The researchers showed that this simple three-neuron block can perform basic brain-like calculations without needing a complex processor.

1. The Comparator (The "Judge"):
Imagine you have two numbers, $V_1$ and $V_2$ . You want to know which is bigger, or if they are both small, or both big.

If $V_1$ is loud and $V_2$ is quiet, the system reacts one way.
If both are loud, it reacts another way.
By looking at which neurons are spiking, you can instantly "classify" the relationship between the two inputs. It's like a judge who raises a specific flag depending on whether the two arguments presented are equal, or if one is much stronger than the other.

2. The Logic Gate (The "Decision Maker"):
They also showed this circuit can act like a basic computer logic gate (the building blocks of all software).

AND Gate: If you set the rules right, the third neuron will only spike if both inputs are high. (1 + 1 = 1).
OR Gate: With a little more "noise" in the system, the third neuron will spike if either input is high. (1 + 0 = 1).

The Real-World Test

The team didn't just run this on a computer; they built it with real hardware.

The Result: The physical device behaved almost exactly like the computer simulation.
The "Messy" Reality: The real-world map was a bit more "fuzzy" than the perfect computer map. You would see a large area where Neuron 1 and 3 spiked, but inside that area, there were tiny "islands" where Neuron 2 would randomly join in.
Why? The researchers realized this fuzziness came from the fact that the metal filaments inside the memristors are constantly changing shape and structure. By adding a "randomness factor" to their computer model to mimic this changing shape, they could perfectly match the messy real-world results.

The Big Picture

This paper claims that by using these "noisy," diffusive memristors, we can create tiny, universal building blocks for future computers. These blocks don't just calculate numbers; they compare signals and make decisions in a way that mimics how biological brains handle information.

The key takeaway is that noise isn't a bug; it's a feature. Just like in a biological brain, the random fluctuations in these electronic neurons help them process information, make comparisons, and perform logic tasks that would be much harder to do with rigid, perfect computer chips.

Technical Summary: Stochastic Dynamics of Diffusive Memristor Blocks for Neuromorphic Computing

Problem Statement
The development of high-energy-efficient artificial intelligence (AI) hardware requires moving beyond the von Neumann architecture toward systems that integrate memory and processing. While diffusive memristors have emerged as promising components for neuromorphic circuits due to their ability to mimic biological synapses (e.g., synaptic plasticity and volatile switching), a critical gap remains in understanding how simple blocks of these devices perform specific computations. Specifically, the mechanisms by which synaptic convergence—the integration of inputs from multiple neurons onto a single downstream neuron—functions within a stochastic, noise-driven environment are not fully characterized. This work addresses the need to identify specific input conditions that enable targeted activation and computational logic in simplified neural circuits composed of diffusive memristors.

Methodology
The study employs a combined approach of experimental fabrication, numerical modeling, and statistical analysis to investigate a three-neuron circuit.

Experimental Setup: The researchers fabricated three identical diffusive memristors (Ag:SiO $_x$ ) using magnetron sputtering and UV photolithography. These devices were configured into a circuit where two pre-synaptic neurons (RM $_1$ and RM $_2$ ) receive external voltage inputs ( $V_1$ and $V_2$ ) and converge onto a post-synaptic neuron (RM $_3$ ). Each memristor is connected in series with a load resistor and in parallel with a capacitor to promote Pearson-Anson oscillator behavior (spiking). The system was subjected to long voltage pulses (10 s ON, 10 s OFF) with input voltages ranging from 0 to 3.6 V. Spiking behavior was classified based on voltage oscillation amplitudes exceeding a 10% threshold above baseline noise.
Numerical Modeling: A charge transport model based on stochastic differential equations was used to simulate the dynamics. This model accounts for the motion of a metallic cluster within an electrochemical potential, driven by electric fields and random thermal fluctuations (modeled via Gaussian white noise). The model incorporates a random parameter $\beta$ to simulate the stochastic reconfiguration of nano-clusters near the filament tip, reflecting the non-stationary nature of filament formation.
Statistical Analysis: To quantify the spiking regimes, the authors calculated the coefficient of variation for inter-spike intervals (CV $_1$ ) and the local variability (CV $_2$ ). These metrics were compared between numerical simulations and experimental data to assess the model's accuracy in reproducing the statistical properties of the spiking patterns.

Key Contributions and Results
The paper establishes a direct link between input voltage combinations and specific spiking patterns in a three-neuron memristive block, demonstrating its potential as a functional unit for neuromorphic computing.

Mapping Spiking Regimes: The study identifies distinct regions on the ( $V_1$ , $V_2$ ) parameter plane corresponding to different spiking outcomes: no spiking, single-neuron spiking (RM $_1$ , RM $_2$ , or RM $_3$ ), dual-neuron spiking, and triple-neuron spiking.
- At low voltages, no spiking occurs.
- As voltages increase, single-neuron regimes emerge.
- Notably, a regime exists where only the post-synaptic neuron (RM $_3$ ) spikes, even when the pre-synaptic neurons do not fire individually. This is attributed to the combined noisy output of the pre-synaptic neurons providing a DC component sufficient to trigger RM $_3$ , or noise-induced spikes occurring when the combined input fluctuates above the threshold.
- Increasing noise levels expands the regions for multi-neuron spiking and can induce spiking even when one input is zero.
Computational Capabilities: The distinct spiking patterns allow the circuit to perform fundamental brain-like computations:
- Classification: The circuit can categorize input pairs ( $V_1$ , $V_2$ ) into different classes based on the resulting spiking pattern.
- Comparison: By fixing one input, the block acts as a comparator, determining if the variable input exceeds specific thresholds.
- Logic Gates: Depending on the noise level and input voltage ranges, the circuit can realize Boolean logic operations. For instance, at lower noise levels, the system functions as an AND gate (outputting a spike in RM $_3$ only when both inputs are high), while at higher noise levels, it functions as an OR gate (outputting a spike if either input is high).
Model-Experiment Agreement: Numerical simulations closely matched experimental measurements regarding the existence of spiking patterns and their statistical characteristics (CV $_1$ and CV $_2$ ). The inclusion of a randomized $\beta$ parameter in the model was crucial for reproducing the "erratic" structure and small "islands" of alternative patterns observed in experiments, which arise from the non-stationary evolution of the conductive filament. The statistical properties of the artificial neurons were found to fall within the range typical of cortical neurons (CV $_1$ and CV $_2$ between 0.2 and 1.5).

Significance and Claims
The authors claim that this work contributes to the development of universal functional blocks for neuromorphic systems. The primary significance lies in demonstrating that a simple, stochastic memristive circuit can perform complex tasks such as signal comparison, classification, and Boolean logic without requiring precise, non-stochastic control.

The paper posits that the inherent volatility and stochasticity of diffusive memristors, often viewed as a challenge, can be harnessed to emulate "neural variability"—a phenomenon in biological systems where neurons respond differently to identical stimuli. This variability is suggested to potentially enhance brain computations. Consequently, the study offers a platform for understanding the computing mechanisms of living neural systems and provides a blueprint for designing large-scale neuromorphic architectures that utilize the natural dynamics of diffusive memristors for information processing. The authors modestly suggest that stabilizing filament evolution could improve reproducibility for specific applications, while the natural volatility remains a promising avenue for emulating biological neural variability.