MSU Researchers Improve Material Property Predictions From Chemical Composition
The new method can estimate the properties of compounds without data on their crystal structure.
Researchers at the MSU AI Center have developed an approach that improves the accuracy of predictive modeling for materials. The study was published in the journal Scientific Reports.
Predicting a material's properties typically requires knowing its crystal structure, or how its atoms are arranged in space. But for many compounds, that structure is not known in advance, and determining it requires substantial computing resources. The researchers therefore focused on predicting properties using only a compound's chemical formula.
To improve accuracy, the researchers transferred knowledge between different types of data. They trained language models using structural information, or first reconstructed a material's structure from its composition and then analyzed it with neural networks. The new models outperformed previous approaches in 25 of 32 tests and reduced prediction error by about 15 percent.
The method could help researchers estimate material properties more accurately when screening promising compounds for further study.
