25-30 August 2019
Henry Ford Building
Europe/Berlin timezone

Deep neural network analysis of DEER data

26 Aug 2019, 12:30
30m
Lecture Hall D (Henry Ford Building)

Lecture Hall D

Henry Ford Building

Talk EPR development and applications EPR

Speaker

Prof. Ilya Kuprov (University of Southampton)

Description

It is demonstrated that deep neural networks (DNN) are a powerful alternative to Tikhonov regularisation methods for the interpretation of double electron-electron resonance (DEER) data. Networks trained using large databases of synthetic DEER traces with carefully modelled distortions and noise are found to process previously unseen experimental data with results comparable to, and occasionally better than, the state-of-the-art Tikhonov methods.

The current best practice for DEER processing is to use Tikhonov regularised deconvolution, a procedure that works well in simple spin-½ pairs, but becomes difficult for more complex systems [1]. Using DNNs trained on simulated data is attractive because training data can be generated to include all complications. Such neural networks can be made resilient to the presence of zero-field splittings, exchange, and out-of-pair inter-electron interactions.

DNN performance strongly depends on the quality of the training dataset. To ensure that the network can successfully process previously unseen datasets, the training database must be representative of the entire range of real experimental systems. The relevant functionality has recently become available in the Spinach library. There are also important factors around the network architecture, pre- and post-processing of data, and the training process.

In this communication we introduce DEERNet and describe how we have created the training database, tuned the network parameters, and handled data pre-processing to achieve excellent performance on real-life DEER data [2].

[1] G. Jeschke, Y. Polyhach, Phys. Chem. Chem. Phys., 2007, 9, 1895-1910.
[2] S.G. Worswick, J.A. Spencer, G. Jeschke, I. Kuprov, Science Advances, 2018, 4 (8), eaat5218.

Primary authors

Mr Steven Worswick (University of Southampton) Mr James Spencer (University of Southampton) Prof. Gunnar Jeschke (ETH Zurich) Prof. Ilya Kuprov (University of Southampton)

Presentation Materials

There are no materials yet.