Researchers at Rice University and Southeast University in China have proposed a material design that could generate true random numbers, the kind used in cryptography, at high speed while preserving the strength of the readout signal.
The study, which Rice University highlighted on September 24, 2026, proposes a mechanism in which magnetization does not flip directly but instead passes through an intermediate state with electric polarization. Theoretical calculations showed that this lowers the energy barrier that must be crossed when switching states, potentially allowing both fast switching and a strong readout signal.
However, this is not a study that measured the performance of a finished cryptographic chip. It is currently a theoretical study based on computer simulations, and it stands as a design guideline for verification in real materials and devices.
Routing magnetization through an intermediate state instead of flipping it directly
The study was carried out by Jun-Jie Zhang, Shuai Dong, and Boris I. Yakobson, and the paper was published on September 9 in the peer-reviewed journal Physical Review Letters.
The subject of the research is a class of materials called "autferroics."
The concept's predecessor was proposed in 2023, and in 2025 a framework using Landau theory was presented to explain its properties. In this study, the researchers examined whether these unusual magnetic and electric properties could be used to speed up true random number generators (TRNGs).
A true random number generator produces random numbers by exploiting unpredictable fluctuations in physical phenomena.
Magnetic approaches use the random flipping of the magnetization of a tiny magnet, caused by thermal fluctuations, and map its two states to 0 and 1.
In a magnetic tunnel junction, electrical resistance changes depending on the direction of magnetization, so 0 and 1 can be distinguished by reading that difference.
The problem is that making magnetization easier to flip tends to weaken the signal used to tell the states apart.
In the conventional magnetic systems the paper uses for comparison, making the magnetic layer smaller or applying a transverse magnetic field makes magnetization flip faster. But the signal difference between the two states also becomes smaller.
No matter how fast the flipping is, it is hard to use as a random number source if the readout circuit cannot reliably distinguish 0 from 1.
Rather than making the flip itself faster, autferroics try to solve this problem by changing the path taken during the flip.
In ordinary multiferroics, magnetism and ferroelectricity coexist within the same phase. In autferroics, by contrast, a magnetic phase and a ferroelectric phase with electric polarization compete strongly and appear as separate states.
As a result, when magnetization changes from one direction to the opposite, it can pass through a state with electric polarization instead of switching directly.
It is like crossing a lower pass to reach the other side rather than climbing straight over a high mountain.
The magnitude of magnetization at the start and end points is preserved, while only the energy barrier that must be crossed along the way is lowered.
This may make it possible to maintain the magnetic signal strength needed for readout even as magnetization switching becomes faster.
About 440 kHz switching and 1.1 MHz random number generation are different figures
The research team ran simulations using TiGeSe3, a monolayer material made of titanium, germanium, and selenium, as an example.
The switching energy barrier, which was 250 meV without magnetoelectric coupling, fell to 87 meV when the coupling was set to 550 meV.
Accordingly, the magnetization switching rate rose from fewer than 100 times per second to about 440,000 times per second, or about 440 kHz.
However, these figures come from a model using an idealized "unit cell," the smallest repeating unit of a crystal structure.
Although the calculations were performed at a temperature of 300 K, this does not mean that an actual TiGeSe3 device delivered the same performance at room temperature.
The authors themselves say the switching rates obtained under these idealized conditions should be regarded as an upper limit.
Also, the speed at which magnetization flips is not the same as the speed at which usable random bit streams are generated.
The researchers applied eight tests from the NIST statistical test suite to bit streams obtained in the simulations, and examined the sampling rates at which all of them could be passed.
The results by condition are summarized below.
| Calculation condition | Metric | Result in the paper | What the figure indicates |
|---|---|---|---|
| Unit-cell model without magnetoelectric coupling | Magnetization switching rate | Below 100 Hz | Baseline switching rate for comparison |
| Unit-cell model with 550 meV magnetoelectric coupling | Magnetization switching rate | About 440 kHz | Switching is faster when passing through an electrically polarized state |
| Same magnetoelectric coupling condition | Random number generation rate passing statistical tests | Above 200 kHz | Simulated rate at which bit streams can actually be extracted |
| Energy difference between magnetic and ferroelectric phases tuned to zero | Random number generation rate passing statistical tests | About 1.1 MHz | Further speedup when the energies of the two phases are aligned |
| Multidomain model including interactions with neighbors | Magnetization switching rate | About 260 kHz | Speed changes when magnetic domains and their boundaries are considered |
The source is Figures 2–4 and the corresponding text of Zhang et al.'s public author manuscript. All values are from simulations, not measurements of physical devices. In the multidomain model, coupling between neighbors was set to 2 meV to examine local switching.
What matters here is that about 440 kHz, above 200 kHz, and about 1.1 MHz each represent a different condition or metric.
About 440 kHz is the speed at which the magnetization itself flips.
Above 200 kHz, on the other hand, is the sampling rate at which the resulting bit streams passed the statistical tests used in the study.
About 1.1 MHz is the value obtained under a further modified condition, in which the energies of the magnetic and ferroelectric phases were brought close to the same level.
As binary random numbers, about 1.1 MHz corresponds to roughly 1.1 million bits per second, but that does not mean 1.1 million bits were extracted directly from 440,000 magnetization flips.
What the research shows is that, besides strengthening the magnetoelectric coupling, reducing the energy difference between the magnetic and ferroelectric phases is also important for fast random number generation.
Changing the switching speed with an electric field while keeping 0 and 1 balanced
Another feature of autferroics is that the switching speed can be tuned by applying an external electric field.
Applying an electric field can change the energy of the ferroelectric state that magnetization passes through during switching.
Importantly, this operation does not directly favor either the positive or negative magnetization direction.
In the team's calculations, even when the electric field changed the switching rate, the proportions in which the two magnetization states appeared stayed at nearly 50-50.
This means it may be possible to build a system in which the random number generation rate can be adjusted without strongly biasing the output toward 0 or 1.
For a random number generator, it matters not only to be fast but also for 0 and 1 to appear without bias.
Four states, not just two, can be used
When the energies of the magnetic and ferroelectric phases are close, it may also be possible to use four states, rather than two, as a random number source.
Magnetization has two directions, positive and negative, and electric polarization also has two directions, positive and negative.
In the research team's model, conditions were shown under which each of these four states appears in about 25% of cases.
If each is assigned
- 00
- 01
- 10
- 11
a single device can yield four-valued random numbers.
However, "four states exist" and "the four appear with equal probability" are different things.
To approach equal probabilities, the energies of the magnetic and ferroelectric phases need to be tuned to similar levels.
The team believes this multistate behavior could also be used for probabilistic computing.
Probabilistic computing does not treat 0 and 1 as fixed values; instead, it represents and computes numbers using the probability that a signal appears.
In this study, the researchers also proposed a configuration that uses the magnetic and electric polarization states to represent the real and imaginary parts of a complex number, and performs operations by combining two junctions with relatively simple logic circuits.
The aim is to simplify processing that would ordinarily require multiple random number sources or circuits by using the multiple states of a single material.
"Quantum-like computation" is not the same as a quantum computer
The paper also touches on using ordinary, probabilistically operating circuits to mimic the probability distributions of quantum mixed states.
This is easy to misunderstand.
What the research shows is a mechanism that uses the multiple states of autferroics to represent probability distributions in hardware.
It does not use qubits, nor does it exploit quantum entanglement.
Therefore, this is not research showing that an ordinary semiconductor chip becomes a quantum computer with this material.
It is more accurate to see it as showing that some probability distributions arising in quantum systems might be efficiently mimicked by conventional probabilistic circuits.
Passing NIST tests does not prove security for cryptographic use
The research team evaluated the random sequences generated in the simulations using eight tests from the NIST statistical test suite.
Passing them is meaningful for confirming that no obvious bias or specific statistical pattern was found in the generated bit streams.
However, NIST has not certified this material or this random number generator.
NIST SP 800-22 itself states that statistical tests alone cannot guarantee that a given random number generator is suitable for cryptographic use, and that they are no substitute for cryptanalysis.
Furthermore, in 2022 NIST indicated a plan to revise this document, explaining that it needs to state more clearly that the SP 800-22 statistical tests should not be used to evaluate cryptographic random number generators themselves.
To judge whether a device is secure enough for uses such as generating cryptographic keys, the statistical properties of the output sequences alone are not sufficient.
It is also necessary to verify whether the physical entropy source is truly unpredictable, whether it can be manipulated from outside, and whether faults or environmental changes could introduce bias.
Practical use first requires demonstrating the material itself
The biggest challenge at present is that the properties of autferroics themselves have not yet been confirmed experimentally.
The present results are based on computer simulations.
The fact that calculations were run at 300 K does not mean that a real device using TiGeSe3 showed performance of about 440 kHz or 1.1 MHz at room temperature.
This study should be positioned as a theoretical design guideline showing what kinds of energy structures and magnetoelectric couplings to look for in materials that could lead to fast true random number generators.
Real materials involve more complex phenomena.
If multiple regions with different magnetization directions, known as magnetic domains, form, the motion of their boundaries, the domain walls, will also affect the switching rate.
When the research team calculated a multidomain model that included interactions with neighbors, the switching rate came to about 260 kHz.
In addition, if defects or structural disorder in real materials make domain walls harder to move, switching could become even slower than in the idealized model.
Numbers obtained for the smallest unit of an ideal crystal cannot be applied as-is to actual, larger devices.
The next step is to confirm experimentally whether candidate materials can truly switch between magnetic and ferroelectric states.
After that, real devices would need to be built and measured for switching speed, power consumption, readout signal strength, quality of the random sequences, and stability over long-term operation.
If these are confirmed experimentally, the study could lead not merely to a material that "switches quickly" but to a way of designing a physical random number source that is fast and can still be read reliably by the circuitry.