Abstract by Yonatan Kurniawan
Physics and Astronomy
Relating Sloppiness and Transferability in Interatomic Models
An Interatomic Model (IM) is a function that estimates the energy of a configuration of atoms. IMs are widely used in material science simulations. Developing an IM can take anywhere from several months to years, relies on expert intuition, and is usually only valid for a single application. Consequently, an important question is how to estimate the transferability of an IM, i.e., how well it will perform on a calculation for which it was not originally intended. One way to understand transferability is through parameter estimation in the IM. The relative importance of the model's parameters can be estimated using the Fisher Information Matrix (FIM). The eigenvalues of the FIM show that many IMs are "sloppy", i.e., have an exponential hierarchy of parameter importance. I discuss how sloppiness quantified by the FIM informs our understanding of transferability.