Thermal microscopy reveals why memristors are inherently random, and that is useful

The conventional picture of a memristor's inner life was simple: apply voltage, a single conductive filament forms through the device. Remove voltage, it dissolves. Repeat. But anyone who has worked with volatile memristors knows the output is noisy, variable, and unpredictable in ways that a single-filament model does not explain. That stochasticity, far from being a nuisance, is exactly what makes memristors attractive for security and computing applications. Until now, nobody could directly observe what produced it.

A team led by Professor Jung Ho Yoon at Sungkyunkwan University, with collaborators at Incheon National University and the Korea Institute of Science and Technology (KIST), has now watched the process happen. Using scanning thermal microscopy (SThM), they measured the Joule heating generated during resistive switching events directly from the top surface of a memristor device. What they saw was not one filament forming and breaking. It was multiple localized hot spots appearing and disappearing as several conductive filaments simultaneously competed for current while ions continuously redistributed inside the device.

Multiple filaments, coupled with thermal effects

The study, published in Advanced Functional Materials, reports for the first time that the resistive switching behavior of ion-motion-mediated volatile memristors originates from a combined mechanism: multiple conductive filaments forming in parallel, coupled with electrothermal effects. The thermal measurements revealed repeated patterns of hot spots at different locations across the device surface, providing physical evidence that the randomness is structural, not merely noise in the measurement.

This multi-filamentary competition explains why memristors produce different outputs even under identical input conditions. Each switching event involves a slightly different configuration of filaments, making the output inherently unpredictable in a way that is very difficult to reproduce or reverse-engineer.

From randomness to encryption

That inherent unpredictability is precisely what cryptographers want. True random number generators (TRNGs) need a physical source of randomness that cannot be predicted even by someone who knows the device's design. Software-based random number generators are, strictly speaking, pseudorandom: their outputs are determined by an algorithm and a seed value. Memristors offer a hardware-based alternative where the randomness emerges from physical processes.

The team demonstrated a bimodal TRNG capable of producing both digital and analog random numbers from the memristor. They then used the generated random numbers as encryption keys to successfully encrypt and decrypt data sequences.

Beyond encryption, the researchers showed the potential for probabilistic computing by performing the inverse operation of a binary full-adder circuit. Probabilistic computing tackles optimization problems that are computationally expensive for conventional deterministic processors by sampling from probability distributions rather than exhaustively searching solution spaces.

What the thermal maps actually show

SThM works by scanning a nanoscale thermal probe across the device surface, detecting tiny temperature variations caused by current flow through the filaments. The key observation was that hot spots appeared at multiple locations simultaneously and shifted positions between switching cycles. This is consistent with multiple filaments forming, competing, and dissolving in a dynamic process governed by ion migration and local heating.

Previous studies of memristors relied on electrical measurements alone, which could detect switching events but could not distinguish between a single thick filament and multiple thin ones. The thermal mapping provides the spatial resolution needed to separate these scenarios.

What the study does not demonstrate

The work is a proof of concept. The TRNG and encryption demonstrations are laboratory-scale, not production-ready systems. Whether the multi-filamentary mechanism can be optimized to maximize randomness quality while maintaining device reliability across billions of switching cycles has not been tested.

The SThM technique, while powerful for research, is not suited to real-time monitoring in deployed devices. It requires laboratory conditions and nanometer-scale probe positioning. The practical value lies in the mechanistic understanding it provides, which can inform device design, rather than in the technique itself becoming part of a product.

Scaling these devices to the density needed for practical computing or security applications also remains a significant engineering challenge. Memristors need to be integrated with conventional electronics, manufactured reliably at scale, and shown to maintain their stochastic properties consistently over time.

But for a community that has been designing memristor devices with an incomplete picture of their internal dynamics, the ability to finally see what is happening inside changes the conversation. Designing for randomness is a lot easier when you understand where it comes from.