Dresden 2020 – wissenschaftliches Programm
Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 64: Topical Session: Data Driven Materials Science - Materials Data Management (joint session MM/CPP)
CPP 64.4: Vortrag
Mittwoch, 18. März 2020, 11:15–11:30, BAR 205
Analysis of Materials Structural Representations for Machine Learning Interatomic Potentials — •Berk Onat1, Christoph Ortner2, and James Kermode1 — 1School of Engineering, University of Warwick, Coventry, United Kingdom — 2Mathematics Institute, University of Warwick, Coventry, United Kingdom
Representations of materials based on atomic structural environments have been used either in machine learning models to predict properties directly or as the core of machine learning interatomic potentials (MLIPs) to enable accurate simulations. Many MLIPs have been developed to translate atomic neighbourhood environments from atom positions to structural representations such as atom-centred symmetry functions, smooth overlap of atomic positions and atomic cluster expansion(ACE) with spherical harmonics. While use of these representations is becoming common practice for applications, the sensitivity of their structural mapping to the materials composition and whether their coverage of the hyper-dimensional space is over-determined or complete have not yet been fully analysed. In this presentation, we provide analysis of the invariance of the model transformation under translations and rotations as well as the sensitivity of descriptors to perturbations. A range of datasets extracted from the NOMAD Archive are used to assess the dimensionality of the representations. The outcomes of our analyses will be presented with discussions on the model sensitivities and their possible limitations. We further provide insights on our continuing affords to utilise structural representations in other models for data-driven materials modelling.