Speaker
Description
Deep neural networks (DNNs) offer a powerful alternative to Tikhonov regularised analysis procedures for standard DEER data. DNNs can be trained on libraries of synthetic DEER traces, allowing them to learn a simulated reality which can include carefully modelled distortions and noise – we have shown that DNNs trained in this manner generalise well to real experimental data. [1]
The well-established Tikhonov methods involve a regularised fitting that relies on a simplified DEER kernel which neglects the exchange interaction (J) completely. In many situations this assumption is reasonable due to the distances and structure between spin-labels. However, the presence of a significant exchange coupling will complicate the interpretation of data under these regimes. It is also often desirable to be able to quantify the exchange coupling, this may present a significant challenge beyond the simpler cases. [2]
In this communication we present new functionality in DEERNet for analysis of DEER in the presence of exchange coupling. By training feed-forward DNNs using data simulated via the full kernel for DEER including exchange coupling, we are able to extract confident estimates for both the inter-spin distance distribution and the exchange interaction.
[1] S. G. Worswick, J. A. Spencer, G. Jeschke, I. Kuprov, Science Advances, 2018, 4, 1-7.
[2] S. Richert et al., Physical Chemistry Chemical Physics, 2017, 19, 16057-16061.