Designing magnets with tunable properties using deep ensembles
(Download Image)
Supercell model of the bilayer containing two interfacial layers. LSCO/LSMO stands for the elements making up the bilayer. These elements are coded as follows—blue: cobalt (Co), magenta: manganese (Mn), green: lanthanum (La), orange: strontium (Sr), and red: oxygen (O).
Perovskite oxides are gaining significant attention for use in next-generation magnetic and ferroelectric devices due to their exceptional charge transport properties and the opportunity to tune the charge, spin, lattice, and orbital degrees of freedom.
Interfaces between perovskite oxides exhibit unconventional magnetic exchange switching behavior, offering a pathway for innovative designs in perovskite oxide-based devices and thereby the design of magnets with tunable properties. Such tunable magnetic materials can have impactful applications for spin-based electronics, sensors, and neuromorphic computers. However, the precise atomic-level stoichiometric compositions and chemophysical properties of these interfaces remain elusive, hindering the establishment of surrogate design principles.
In a study published in ACS Applied Materials & Interfaces, LLNL researchers and a collaborator from UC Davis designed deep learning model ensembles to investigate the effects of composition and process parameters on the magnetic behavior of perovskite oxide multilayers made up of lanthanum (La), strontium (Sr), cobalt (Co), oxygen (O), and manganese (Mn). The deep learning approach enabled the team to search through over 50,000 interface structures for the most stable ones, while accurately describing their magnetic properties. Their model elucidated the role of strontium segregation and oxygen depletion on opposite sides of interfaces in the multilayers, such that functional changes in the structure, charge transfer, and magnetic properties of the system can be controlled by varying strontium doping and layer thicknesses.
The model leveraged first-principles simulations using density functional theory (DFT), evolutionary algorithms, and neural network searches with on-the-fly uncertainty quantification. Many interesting properties can be overlooked with DFT calculations alone, as DFT calculations for systems containing lanthanides and transition metals are resource-intensive and computationally challenging, limiting the number of atoms that can be included and the total number of systems to be considered. Neural network-based deep learning models, on the other hand, exhibit remarkable flexibility and excellent scalability and are computationally more efficient compared to brute force DFT calculations—which were essential in this work to explore the full compositional phase space.
Combining these methods allowed researchers to screen materials with fewer computing resources, analyzing the interfaces in the multilayers and their properties with unprecedented levels of sophistication. This research opens new avenues for efficiently screening multicomponent materials with tunable properties for functional applications.
This work was funded by the Laboratory Directed Research and Development (LDRD) Program at LLNL (project tracking code 21-ERD-005).
[H. Sun, V. Lordi, Y. Takamura, A.Samanta, Unraveling the Correlation between the Interface Structures and Tunable Magnetic Properties of La1–xSrxCoO3−δ/La1–xSrxMnO3−δ Bilayers Using Deep Learning Models, ACS Appl. Mater. Interfaces (2024), doi: 10.1021/acsami.3c18773.]
