Brown Engineers' Breakthrough Could Enable Energy-Efficient AI Chips

Brown University researchers developed a technique to make ultra-thin multiferroic materials that retain their functional properties, offering a potential path toward low-power microchips for AI.

PROVIDENCE, R.I. [Brown University] - Multiferroics are a promising class of materials for making computer chips and other electronic components that can operate at much lower power than today's devices. The materials have coupled electrical and magnetic properties that can be manipulated simultaneously with a relatively tiny electrical voltage.

Engineers are hopeful that multiferroics could help reduce the energy burden associated with the rise of AI, but there's a problem known as the "dead layer" effect: When these materials are made thin enough to use in a computer chip, they lose the electrical and magnetic properties that make them so interesting.

Now, a team of Brown University researchers has developed a technique for making thin films of the multiferroic bismuth ferrite that retain their key properties and have some emergent properties that make them even more useful.

"We've established that we can get rid of the dead layer," said Lucas Caretta, an assistant professor of engineering at Brown who led the work. "We show that you can make this material basically as thin as you want - down to a few angstroms - and it's still ferroelectric, and it's still magnetic."

The research is published in Nature Materials.

The interest in multiferroic materials stems from their linked electrical and magnetic properties. Each unit cell of the material's atomic lattice is electrically polarized, meaning positive and negative charges align on opposite sides of the cell. That electrical polarization can be flipped - meaning the positive and negative charges switch sides - by applying a small voltage to the cell. That means it's possible to encode binary data (ones and zeros) by manipulating polarization states, which can be done through low-power voltage pulses. And because the magnetic states of the material are linked to the electrical states, the data can be read back out magnetically, creating an ultra-low-power memory or logic device.

The dead layer phenomenon, however, has historically been a deal-breaker. To be integrated into a computer chip, a multiferroic material must be made exceptionally thin. But at low thicknesses, two problems begin to arise: The material's atomic lattice loses the physical shape and chemistry it needs to function, and internal electric fields overpower its electrical polarization. Together, these twin breakdowns cause the material to lose both its ferroelectric and magnetic properties entirely, hence the so-called dead layer.

The Brown team's solution is a new way of fabricating bismuth ferrite, a widely studied multiferroic. The researchers grew the bismuth ferrite film on a substrate made from dysprosium scandate, a rare Earth-derived material with a crystalline structure that's narrower than that of bismuth ferrite. That lattice mismatch creates a compressive strain on the bismuth ferrite, which helps to maintain its functional atomic lattice structure. Meanwhile, a second metallic layer helps to take up stray electric fields, which maintains electrical polarization in the bismuth ferrite layer. The result is a bismuth ferrite film that not only retains its functional multiferroic properties, but actually has enhanced magnetic properties.

The researchers showed that the stacked structure causes the bismuth layer to become an altermagnet - a kind of hybrid between a ferromagnet and an antiferromagnet.

"Ferromagnets are useful for memory and logic because they produce strong magnetic signals that are easy to read out," Caretta said. "The challenge is that they are relatively slow, and their stray magnetic fields make it difficult to pack devices densely on a chip without interference. Antiferromagnets can operate much faster and avoid those stray fields, but their signals are much harder to detect. Altermagnets are kind of the best of both worlds - they can be fast, densely integrated and still provide a strong electrical signal for readout."

Caretta says he's hopeful that the results are a step toward bringing multiferroics into play as components in low-power computing.

"We know the underlying properties are already there in the bulk materials," Caretta said. "The trick is learning how to engineer them into a form that's useful - and sometimes even better than what nature gives us."

The research was supported by the U.S. Army Research Office (W911NF-21-2-0162), the National Science Foundation (DMR-2039380, OIA-2327352) and the U.S. Department of Energy (DE-AC02-05CH11231).

/Public Release. This material from the originating organization/author(s) might be of the point-in-time nature, and edited for clarity, style and length. Mirage.News does not take institutional positions or sides, and all views, positions, and conclusions expressed herein are solely those of the author(s).View in full here.