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 103: Topical Session: Data Driven Materials Science - Machine Learning for Materials Characterization (joint session MM/CPP)

CPP 103.3: Vortrag

Donnerstag, 19. März 2020, 16:30–16:45, BAR 205

Bayesian models and machine-learning for NMR crystal structure determinations — •Edgar Albert Engel¹, Andrea Anelli², Albert Hofstetter³, Federico Maria Paruzzo³, Lyndon Emsley³, and Michele Ceriotti² — ¹TCM, University of Cambridge, United Kingdom — ²COSMO, Ecole Polytechnique Federale de Lausanne, Switzerland — ³LRM, Ecole Polytechnique Federale de Lausanne, Switzerland

NMR spectroscopy is a key tool for determining the atomic structure of powdered and amorphous solids, which usually proceeds by finding the best match between experimentally observed NMR chemical shifts and those of candidate structures. However, the reliability of structure determinations depends on the errors in the predicted shifts. I will demonstrate how a Bayesian approach based on knowledge of the typical errors, coupled to visualisations of the similarity of the candidate structures, allows to quantify and understand the resultant confidence in the identifications of the experimental structure [1]. The applications highlight that using self-consistently determined uncertainties instead of commonly used global estimates make it possible to use 13C shifts to improve the accuracy of structure determinations. I will further outline how a machine-learning approach including uncertainty estimation [1,2] ties in with the above structure determination framework and that it can provide a surrogate for costly or even outright unfeasible first-principles predictions of NMR shifts.

[1] E. A. Engel et al., Phys. Chem. Chem. Phys., 21, 23385 (2019) [2] F. M. Paruzzo et al., Nature Comm., 9, 4501 (2018)