A materials breakthrough at Brown University could give future AI hardware a dramatically more efficient way to process information. Engineers have developed a method for shrinking films of bismuth ferrite to nearly atomic dimensions without destroying the electrical and magnetic behavior that makes the material useful. The advance targets a major obstacle to ultra-low-power chips at a time when AI systems are driving demand for ever more electricity.
Bismuth ferrite belongs to a class of materials called multiferroics, which can combine ferroelectric and magnetic properties. That combination is attractive for memory and logic hardware because it could allow information to be manipulated through multiple physical mechanisms instead of relying exclusively on conventional charge-based switching. The catch has been thickness: once these materials are reduced far enough for practical chip architectures, their useful properties tend to collapse into what researchers call a “dead layer.”
Brown’s team, led by engineering assistant professor Lucas Caretta, addressed both sides of that failure. The researchers grew the film on dysprosium scandate, a substrate whose narrower crystal lattice places compressive strain on the bismuth ferrite and helps preserve its atomic structure. A metallic layer in the stack also absorbs stray electric fields, protecting the film’s electrical polarization. Brown says the resulting films remain ferroelectric and magnetic at thicknesses of only a few angstroms. The university’s account of the work says the same structure also produces enhanced magnetic behavior.
That enhanced behavior may be as important as the removal of the dead layer itself. The stacked material takes on an altermagnetic state, a comparatively new magnetic arrangement that combines features associated with ferromagnets and antiferromagnets. Such states could offer researchers additional ways to control spin and encode information, potentially expanding the design options for spintronic devices. The MAX IV laboratory, which supported the materials characterization, likewise describes the technique as a possible route toward lower-power microchips for AI.
The result is not an instantly deployable AI processor, and the work still has to move from carefully controlled thin-film experiments toward repeatable device fabrication and circuit-level testing. Even so, solving the thickness problem removes one of the biggest barriers facing multiferroic electronics. If engineers can integrate these films with existing manufacturing processes, the approach could eventually help reduce the energy cost of moving and storing data inside AI hardware—one of the field’s most stubborn engineering challenges.
