Dresden 2020 – scientific programme
The DPG Spring Meeting in Dresden had to be cancelled! Read more ...
Parts | Days | Selection | Search | Updates | Downloads | Help
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 103: Topical Session: Data Driven Materials Science - Machine Learning for Materials Characterization (joint session MM/CPP)
CPP 103.4: Talk
Thursday, March 19, 2020, 16:45–17:00, BAR 205
Teaching machines to learn dynamics in NMR observables — •Arobendo Mondal, Karsten Reuter, and Christoph Scheurer — Chair for Theoretical Chemistry, TU Munich, Germany
NMR is a powerful tool for studying the structural and electronic properties of molecules and solids. However, the interpretation of NMR spectra for large systems is often challenging as a result of the free or constrained dynamics of the ligands attached to the NMR active nucleus. Computed NMR parameters can aid in the interpretation. Their accuracy depends on the level of method used, with the high computational cost of highly accurate first-principles calculations quickly limiting the tractable system sizes and number of such computations.
In this respect, emerging machine learning approaches are an appealing option. The key challenge here is an efficient data representation, as NMR parameters depend strongly on their local chemical environment with often non-negligible effects of the second and third coordination sphere. To this end, we use a combination of multiple SOAP descriptors1 to learn NMR parameters for the Antamanide peptide molecule from quantum chemical data computed on a small subset of a long 90 ns molecular dynamics trajectory. The trained model is found to predict NMR parameters within DFT accuracy for 90,000 snapshots from this trajectory that were not contained in the training data.
[1] Phy. Rev. B, 2013, 87, 184115