2022 | 2021 | 2020 | 2019 | 2018 | 2017 | 2016 | 2015 | 2014 | 2013 | 2012 | 2011 | 2010 | 2009 | 2008 | 2007 | 2006 | 2005 | 2004 | 2003 | 2002 | 2001 | 2000 | 1999 | 1997

Machine Learning Prediction of the Stability of Perovskites

Authors: J. Schmidt

Ref.: Bachelor Thesis, Martin-Luther University of Halle-Wittenberg (2016)

Abstract: Often technical solutions for humanity’s problems already exist in theory, the only problem being that either the materials needed to implement the solutions are too toxic, rare or expensive to use or they do not exist yet. Every new material that is discovered offers a minuscule chance to solve one of these problems. Unfortunately, the classic way of finding such new compounds is extremely slow. Classically experimental physicists or chemists start from known structures and try to synthesize similar materials without concrete knowledge on the stability of the new material.
New theoretical methods like Density Functional theory (DFT), better algorithms and the exponential rise of computing power paved the way for theoretical structure prediction methods. Nowadays, by using DFT, one can predict the stability of a new compound by calculating the formation energy of the material and all competing phases of the compound and finding the global minimum. Even though this process is orders of magnitude faster than classical methods it is still slow, because DFT calculations and finding minima of a high-dimensional surface are computationally expensive and need millions of hours of CPU time.
If one could somehow presort the set of potential new materials to reduce its size without excluding stable materials, one could greatly accelerate the process of discovering new materials. In order to do this one has to build a different stability prediction model that is faster but not as accurate as DFT.
The idea behind this thesis to train and tests machine learning models on existing DFT-data for perovskites and evaluate their stability-prediction abilities.