Many methods in computational materials science require scientists to identify an appropriate set of parameters that capture the physics of the particular material being studied. It may be possible to calculate these parameters from scratch, but it takes a lot of time and computing power. Therefore, scientists are always keen to find more efficient ways to estimate them without performing full calculations.
This is the case for Koopmann functionals. The Koopmann functional is a promising approach that extends the power of density functional theory and allows it to be used to predict not only the ground state but also the spectral properties of materials, such as the frequencies of light that they absorb. (such as the optimal location of atoms within that material). The accuracy of the Koopman function depends on finding appropriate “screening parameters” for the system under study.
“The screening parameter can be interpreted as a measure of how responsive the remaining electrons in the system are to the addition or removal of electrons,” said Edward Linscott, a postdoctoral fellow at the Pall Center for Scientific Computing Theory Data. explain. Member of the Scherer Institute and MARVEL.
The term “screening” refers to the fact that other electrons mask, or in other words, shield, new electrons from being added to someone monitoring the system from the outside.
Linscott continued, “And this electron screening, the process of adding or removing electrons from a system, is the physical process that we are interested in when discussing spectral properties. For example, the solar By shining light on the cell’s photovoltaic material, it releases electrons, creating an electrical current.
Using machine learning
The disadvantage of Koopman’s function calculation is that it is more time-consuming than DFT’s function calculation, mainly due to the computational cost of the screening parameters. However, a new paper published in npj Computational Materials shows that even simple machine learning models trained on modest amounts of data can significantly reduce the time required to calculate the screening parameters in the Koopman algorithm. It shows.
This article was co-authored by Yannick Schubert from the University of Zurich (who started this project as a master’s thesis), Sandra Luber (Yannick’s PhD supervisor at UZH), Nicola Marzari, Director of MARVEL, and Lin. Written by Scott himself.
The authors selected two specific materials for their study: liquid water and the halide perovskite CsSnI3.
“These represent the kinds of systems that we thought could make the most of machine learning,” Linscott explains. “Liquid water is naturally disordered and out of the comfort zone for us scientists who have worked with pristine crystals. Halide perovskites, on the other hand, are promising materials for use in solar cells and It is important to calculate how it changes.
Modeling the spectral properties of these systems using Koopman functionals requires calculations on many copies of the same chemical system with different atomic positions. This process can be made much faster by training a machine learning model on a subset of these copies and using this model to predict screening parameters for the remaining copies.
When the researchers embarked on their study, they wondered how machine learning models would work among the many options offered in this field. It turns out that a simple model called ridge regression is sufficient.
“We had a roadmap for increasingly sophisticated networks that we were going to try, but we didn’t know from the beginning whether it would work or whether it would generate enough data. “I did,” Linscott said.
“Surprisingly, the simplest model with very little data worked well. The model we ended up with is completely different from the sophisticated machine learning models that are all around us these days. was sufficient to accurately calculate the screening parameters.”
While the network itself is simple, scientists say this is partly due to their focus on building “descriptors,” which are mathematical objects that encapsulate the relevant physical phenomena of a system and need to be input into machine learning models. I believe this is due to the careful work that was done.
While it is possible to extend this method and make it more powerful with more complex networks and more training data, Linscott said the next step is to take full advantage of the current method to It states that it is to be used for research. -Depends on the spectral properties of the interesting material.
