Sign In

Communications of the ACM

ACM TechNews

ML Method Makes Better Predictions of Material Properties Using Low Quality Data


View as: Print Mobile App Share: Send by email Share on reddit Share on StumbleUpon Share on Hacker News Share on Tweeter Share on Facebook
Logo of the University of California, San Diego Materials Virtual Lab.

Nanoengineers in the University of California, San Diego Materials Virtual Lab developed a machine learning method to predict the properties of materials with greater accuracy than existing models.

Credit: University of California, San Diego

Nanoengineers at the University of California, San Diego (UCSD) have developed a new machine learning (ML) method that uses low-quality data to predict material properties with greater accuracy than existing models.

UCSD's Shyue Ping Ong said the method works to "combine the large low-fidelity data and the smaller high-fidelity data to improve the models' accuracy in high value predictions."

The UCSD team examined the materials' band gaps to determine their electrical conductivity and other properties as a proof of concept; their multi-fidelity graph networks led to a 22% to 45% reduction in the mean absolute errors of experimental band-gap predictions, versus a single-fidelity model.

The approach also accurately forecast high-fidelity molecular energies, and could generate a predictive model for disordered compounds.

Said Ong, "Now we are able to do materials discovery and prediction across the entire space of both ordered and disordered materials, rather than just ordered materials."

From UC San Diego News Center
View Full Article

 

Abstracts Copyright © 2021 SmithBucklin, Washington, DC, USA


 

No entries found