CaDS Seminar 2022 - Aug. 30

Xinzhe Wu (SDL Quantum Materials)

Learning Explicit Functions for the Excess Enthalpy of mixing  of LnPO4 by Sparse Regression

Abstract:
In recent years, the role of existing Machine Learning (ML) methods have experienced a tremendous growth in many scientific domains including Materials Science and Quantum Chemistry. However, for a lot scientific applications, there are still gaps between the ML methods and the requirements of scientists. Unlike most applications of machine learning, which learn and predict in a black box mode, the target of a lot of scientific applications is to learn an interpretable model which is acceptable in the sense of science. This is one of the difficulties which bring ML methods to the scientific community.
In this talk, I will present our attempt, as SDL Quantum Materials, on this topics based a specific chemical application. This is a mini-project, which tries to identify the most relevant descriptors for the excess enthalpy of mixing for Lanthanide Orthophosphate (LnPO4) in two types of phosphate minerals. This is a typical regression problem, with further constraints: 1) explicit functional formulas should be learnt; 2) they can be interpreted by physicists; 3) predicting the excess enthalpy in acceptable accuracy. In this work, we try to build a series of explicit functions for the excess enthalpy of mixing based on the combinations of sparse regression and Kernel Ridge regression.