Physics-Informed Machine-Learning Prediction of Curie Temperatures and Its Promise for Guiding the Discovery of Functional Magnetic Materials

Abstract

High-performance permanent magnets with a high Curie temperature, containing less critical materials, are integral to zero-carbon energy solutions. We built a machine-learning model trained over available experimentally measured Curie temperature values to predict the T C of multicomponent magnetic materials. We chose two compositions from a pseudo-binary (Zr 1−x Ce x )Fe 2 system, namely, (Zr 0.16 Ce 0.84 )Fe 2 and (Zr 0.94 Ce 0.06 )Fe 2 , to experimentally validate the ability of our model to predict the Curie temperature of novel compounds. We also provided a detailed discussion on the correlation of the Curie temperature with the de Gennes scaling factor in rare-earth intermetallic compounds and its breakdown below a certain rare-earth content. The electronic structure calculations (density of states and Fermi surface) were performed using the density functional theory on selected compounds (Zr 0.16 Ce 0.84 )Fe 2 and (Zr 0.94 Ce 0.06 )Fe 2 to understand the electronic origin of a strong magnetic exchange. We found that the change in the electronic density of states and electron/hole fillings at the Fermi level directly correlate with the Curie temperature. Notably, our model was able to capture these key electronic structure trends, which show that physics-informed machine learning can play a crucial role in designing new high-performance magnets with improved properties for environmentally sustainable applications.

Highlights from 2023 and Chemistry of Materials’ 35th Year

“…The manuscripts that conclude our list are diverse and include a report of physics-informed machine-learning for prediction of Curie temperatures of magnetic materials, complex perovskite oxides for solar thermochemical watersplitting, and conductive polymers containing indacenodithiophene moieties for thermoelectric applications . Kim and co-workers also have a highly engaged with manuscript on conformal growth of hexagonal boron nitride, which aligns well with the journal’s interest in Precision Patterning that will be featured in 2024 in a special collection of invited and submitted content.…”

MagBERT: Magnetics Knowledge Aware Language Model Coupled with a Question Answering Pipeline for Curie Temperature Extraction Task

“…This can be performed only for select materials let alone in a high throughput manner due to its prohibitive computational complexity . Applying machine learning methods for the materials containing 4f electrons are also limited to specific class of compounds due to the limited amount of data present . For example, Ghosh et al, established structure–property links in uranium and neptunium based compounds by developing machine learning models and found that cation f-subshell occupation numbers was the key descriptor in predicting the magnetic moments in these structures .…”

Machine-Learning Prediction of Curie Temperature from Chemical Compositions of Ferromagnetic Materials

“…In 3-fold cross-validation, an MAE of 73 K with a standard deviation of 3.2 K was achieved using the first data set, while an MAE of 71 K with a standard deviation of 2.3 K was achieved on the combined data sets. Another study using a random-forest regression was conducted by Singh et al, 28 which attained a 5-fold cross-validation R 2 of 0.91 with a root-mean-square error (RMSE) of 59 K. Nevertheless, this study used a relatively small data set of 220 ferromagnetic and ferrimagnetic compounds to analyze rare-earth-based materials. Moreover, a linear regression was used by Sanvito et al 29 in order to accelerate the discovery of new ferromagnets in the Heusler alloy family.…”