Session ThOD. There are 4 abstracts in this session.

Session: Theory & Simulation, time: 4:00-4:25
Deep neural network processing of DEER data
Steven Worswick1; James Spencer1; Gunnar Jeschke2; Ilya Kuprov1
1University of Southampton, Southampton, United Kingdom; 2ETH Zurich, Zurich, Switzerland
Double electron-electron resonance (DEER) is an essential EPR tool, used to measure distances in paramagnetic or paramagnetically tagged biomolecules. DEER spectroscopy involves recording a dipolar modulation signal between two unpaired electrons and running a Tikhonov regularised fitting to extract the distance distribution. The procedure runs into difficulties when more than two unpaired electrons are present, or when their spin exceeds 1/2. Here we describe an attempt to process DEER data using neural networks trained on large databases of simulated data. Accuracy and reliability of neural network outputs from experimental data was found to be unexpectedly high. The networks are also able to identify ambiguous and corrupt datasets.

Session: Theory & Simulation, time: 4:25-4:50
Solving Structures with Weighted Dice: Routine Application of Quantum Chemistry to NMR Problems
R. Thomas Williamson
Merck & Co., Inc., Rahway, NJ
Theoretical calculations of molecular geometries and NMR parameters have advanced in recent times. These calculations provide a framework for the design of novel NMR experiments and revitalization of older experiments. Applications important within the pharmaceutical industry include calculations to support studies aimed at defining the constitution and configuration of small molecules, analysis of peptides in anisotropic media, conformational sampling, and spectral simulations for the analysis of crystal structures. Theoretical methods based on QM DFT calculations provide better strategies for choosing the optimal experiment for a specific task (e.g. using HMBC vs. ADEQUATE). It is now common to simultaneously utilize two or three orthogonal methods simultaneously (e.g. RDC/RCSA, chemical shifts, J-couplings, NOE/ROE, etc.) to afford robust structure confirmation.

Session: Theory & Simulation, time: 4:50-5:05
Predicting 1H and 13C Chemical Shifts of Molecular Crystals with DFT Accuracy Using Machine Learning
Federico Maria Paruzzo; Albert Hofstetter; FĂ©lix Musil; De Sandip; Michele Ceriotti; Lyndon Emsley
EPFL, Lausanne, Switzerland
Chemical shift based solid-state NMR approaches are emerging as powerful methods for structure elucidation of amorphous materials and crystalline solids. However, the approaches suffer from high computational costs associated with the required chemical shift calculations. We present a machine learning model based on Gaussian Process Regression and local SOAP fingerprints, which was recently shown to achieve density functional theory (DFT) accuracy for energy and force calculations. The model is trained on the DFT calculated chemical shifts of a set of molecular crystals. The prediction performance is demonstrated to achieve accuracy comparable with DFT for a set of 500 randomly selected molecular crystals, while reducing the calculation time from hours to seconds per structure.

Session: Theory & Simulation, time: 5:05-5:20
SIMPAIN - SIMPAIN - Simulation Package for Analytical Insight  into complex pulse sequences in Nuclear Magnetic Resonance 
Anders Bodholt Nielsen
Interdisciplinary Nanoscience Center (iNANO), Aarhus C, Denmark
We introduce a new simulation package to analyze complex pulse sequences in NMR. The program can be employed to approximate an effective time-independent Hamiltonian under static or magic-angle-spinning (MAS) conditions. We show that by transforming each individually spin into a radio-frequency (rf) interaction frame, we can calculate the effective Hamiltonian up to 2. order. As examples of pulse sequence, we analyze TOCSY experiments in liquid-state NMR and the effect of anisotropic chemical shift (CSA) under MAS.