Machine Learning Models of Solid Properties for High-Throughput Screening of Condensed Phase Materials with Chemical Accuracy
Abstract
We have investigated several machine learning models which can be used to study solids. First findings were published in a peer reviewed journal article. Within a second study, we developed a new representation, and applied it to the study of elpasolite crystals. Elpasolite was chosen since it is the predominant quaternary crystal structure (AlNaK2F6 prototype) reported in the Inorganic Crystal Structure Database. We generated machine learning model to calculate density functional theory quality formation energies of all the 2M pristine ABC2D6 elpasolite crystals which can be made up from main-group elements (up to bismuth). Our model's accuracy was improved systematically, reaching 0.1 eV/atom for a training set consisting of 10 k crystals.Important bonding trends are revealed and are reported. Subsequently, in 2016 and 2017, a more universal and improved representation was developed and tested on various data sets including molecules, solids, water clusters, and peptide side chain interactions. Systematic improvement of prediction error with training set size was demonstrated for all data sets, often reaching unprecedented predictive power. Impressive results were also obtained for various molecular electronic ground-state properties,dipole moments, HOMO-LUMO gaps, and polarizabilities. Interestingly, learning has even been observed when predicting systems containing chemical elements which were absent in training. The results of these studies were published in an article in 2018.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 20, 2018
- Accession Number
- AD1059550
Entities
People
- Otto A Von Lilienfeld-toal