Updated machine learning descriptors allow faster and more accurate predictions of new materials

What is it about?

Computers can design and predict new materials that make research and development faster and cheaper. One way to predict the properties of a new material is through machine learning potentials (MLPs). A key component of MLPs is a description of the local atomic environment. Special math terms called “descriptors” are a key part of this description. These descriptors need to meet certain properties to allow the MLP to function properly. This paper sets out an updated version of a type of descriptors known as “Spherical Bessel” (SB) descriptors. It shows that the updated SB descriptors satisfy certain important properties and help make faster calculations.

Featured Image

Photo by Antoine Dautry on Unsplash

Photo by Antoine Dautry on Unsplash

Why is it important?

Descriptors need to fulfill certain math conditions to make accurate predictions. The first condition is that descriptors should not be affected by changes to the symmetry of the system. Another condition is that they should be 'twice-differentiable with respect to atomic positions.' This ensures that the forces described by the descriptors during calculations are continuous. It also ensures that the descriptors can be computed efficiently. The final property is “completeness." This allows the local atomic environment to be rebuilt up to a desired level. A stronger version of this condition is “optimal completeness.” This condition means that completeness must be satisfied with the fewest possible number of descriptors. So far, none of the descriptors described in literature satisfy all these conditions. The updated SB descriptors described in this study meet the first two conditions. They also meet an important requirement for optimal completeness. The study also compared the updated SB descriptors to commonly used descriptors. It showed that the updated SB descriptors permit calculations that are nearly 10 times faster. KEY TAKEAWAY: Descriptors are important math terms for the prediction of new materials using MLPs. The updated SB descriptors described here satisfy some important properties for descriptors. They also allow nearly 10 times faster calculations than other commonly used descriptors. Keywords/meta tags: