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

Active Learning the Thermodynamic Stability of Solids

Authors: Matteo Tabusso

Ref.: Bachelor thesis, Martin-Luther University of Halle-Wittenberg (2020)

Abstract: Using a set of 64 elements (hydrogen to bismuth, excluding lanthanoids and noble gases) we arrive at 249.984 possibilities for a ternary compound. By increasing the elements per compound to four, over 15.2 million possibilities emerge. This is practicably impossible to solve for quaternary compounds as already the 250.000 possibilities of ternary compounds take hundreds of thousand CPU-hours to compute. This is where the main advantage of ML takes effect. A properly trained model could possibly predict any materials stability and characteristics with sufficient accuracy within seconds. Only the training of the model and obtaining sufficient training data costs time but these two points should still be multiple orders of magnitude faster than the standard high-throughput approach. However, if we consider a quaternary or quintary composition space even a small training set of a few percents of the total space might be too expensive in practice. In this thesis we will research ways to reduce the size of the training set through an active learning approach, which allows a target selection of the most important compounds for training.

Download