AI-Driven Discovery: Two Room-Temperature Magnetic Semiconductor Candidates

Announcement: Vals AI Research is unveiling two potential magnetic materials for the next generation of computer memory, identified by a collective of Claude Opus 5.5 agents.

These candidates are predicted to maintain zero net magnetism while still possessing the ability to sort electrons by their spin. One is a completely original compound engineered by the AI, while the other is a material that was first synthesized back in 1999.

To ensure transparency, we are providing the full suite of calculations, the underlying code, and a detailed list of known caveats.

Author: Geby Jaff | Date: 10/04/2026

Most of us are familiar with two primary categories of magnets. The first is the ferromagnet (like a standard fridge magnet), where atomic magnets align in the same direction (all $\uparrow$ or all $\downarrow$ ), compounding their magnetic strength. The second is the antiferromagnet (AF), where neighboring atomic magnets point in opposite directions, effectively canceling each other out.

For years, computer memory researchers have been striving to find materials that bridge the gap between these two extremes. To understand why, we need a clear conceptual framework.

A team of AI agents and I have identified two candidates—one designed from scratch and one rediscovered from 1999—that our simulations suggest possess the ideal properties. For those who aren't physics or chemistry undergraduates, here is a ~~complex~~ rapid 90-second primer on magnetism.

🧲 A 90-Second Magnet Primer

The Concept of Spin

Every electron possesses a quantum mechanical property known as spin, which generates its magnetic moment. We can simplify this by modeling the spin direction as either up or down.

In the field of spintronics, information is stored based on this orientation. Examples include:

Hard drive read heads: Devices that interpret magnetic bits on a platter.
MRAM: Non-volatile memory that stores binary data using spin principles.

The goal of spintronics is to sort electrons by spin to read or write data. While ferromagnets do this naturally, standard antiferromagnets cannot.

Comparing Three Magnetic Types

1. Ferromagnets (Ferro)

These materials produce a macroscopic magnetic field that leaks from the surface (which is why they stick to fridges). However, this creates issues:

~~Easy to control~~ $\rightarrow$ The field interferes with nearby components and is difficult to isolate.
Switching states is power-hungry and relatively slow.
Benefit: Spins are sorted by energy level; at the edges of the energy spectrum, electrons share the same spin.

2. Antiferromagnets (Antiferro)

These are the inverse of ferromagnets: spins are unsorted.

Benefit: No macroscopic field allows for much denser packing of storage components.
Benefit: Switching speeds are roughly $1,000\times$ faster than ferromagnets.
Drawback: Because electrons of the same energy level have mixed spins, reading/storing spintronic information is nearly impossible.

3. Luttinger Compensated (LC)

LC materials are a specialized subset of antiferromagnets. They maintain a net spin moment of zero: $\sum \text{spin} = 0$ However, the spin-up and spin-down atoms exist in inequivalent environments. This happens if:

They are different chemical elements (e.g., Element A $\uparrow$ and Element B $\downarrow$ ).
They are the same element but occupy different types of crystal sites.

According to Luttinger’s theorem, in an insulator, the net spin moment of each repeating unit must be an integer; thus, once it hits zero, it remains locked. While heat and spin-orbit coupling can cause slight imbalances, the inequivalence of the sites allows spins to be sorted by energy, similar to ferromagnets.

The "Spin Window" and Thermal Noise

The critical factor for memory storage is the spin window: the energy slice at the edge of the band gap where all available electron states share the same spin.

To be viable at room temperature, this window must be larger than the thermal energy (the "jiggling" of atoms), which is approximately: $\text{Thermal Energy} \approx 26 \text{ meV}$

The Ideal Goal: A semiconductor with a band gap, zero net magnetism, and a spin window significantly larger than $26 \text{ meV}$ .

The Discovery Process

The AI agents utilized Density Functional Theory (DFT) to run quantum-mechanical simulations. They employed two levels of approximation to balance speed and precision:

Method	Speed	Accuracy	Use Case
`PBE+U`	Fast	Moderate	Initial Screening
`HSE06`	Slow	High	Final Validation

The following results are derived from the more precise HSE06 calculations.

Candidate 1: The Designed LC Magnet ( $YBaMnFeO_5$ )

The agents engineered a brand-new compound consisting of these five elements:

As far as we could find, it h...

# Conceptual representation of the DFT workflow used by agents
def screen_material(material):
    if run_simulation(material, method="PBE+U").is_promising():
        result = run_simulation(material, method="HSE06")
        if result.spin_window > 26: # meV
            return "Candidate Found"
    return "Discard"