ShiftML : chemical shifts in molecular solids by machine learning

ShiftML was developed jointly by the Ceriotti and Emsley groups at the EPFL [1]. ShiftML uses a machine learning framework to predict chemical shifts in solids which is based on capturing the local environments of individual atoms. ShiftML has been trained on 3564 structures taken from the Cambridge Structural Database (CSD), chosen to be as diverse as possible. On a separate test set of 604 randomly selected structures it predicts chemical shifts of molecular solids to an accuracy with respect to GIPAW DFT of (RMSE) 0.48 ppm for ¹H, 4.13 ppm for ¹³C, 13.7 ppm for ¹⁵N, and 17.05 ppm for ¹⁷O and with an R² of 0.97 for ¹H, 0.99 for ¹³C, 0.99 for ¹⁵N, and 0.99 for ¹⁷O. The machine learning model also reports a measure of confidence in the prediction, in the form an uncertainty on the predicted shifts, based on the framework introduced in [2].

ShiftML takes as an input a crystal structure (in a number of different formats), and

• predicts the isotropic NMR chemical shieldings/shifts of ¹H, ¹³C, ¹⁵N and ¹⁷O;
• provides a JSmol interactive output, and copy-paste content in the extended xyz format, in the magres format (the calculated shieldings/shifts are given as a tensor with the isotropic shielding/shift as all three diagonal elements) and in JSON.

ShiftML has so far been tested for DFT-optimized crystal structures of molecular solids containing ¹H, ¹³C, ¹⁵N, ¹⁷O and ³³S atoms (although ³³S chemical shieldings are not predicted, as they are very rarely used in practice).

Version 1.1; 19.07.2019

This is the first point release of ShiftML, yet it brings substantial changes to app. This version has been built on a new model and allows prediction of chemical shifts of structures containing also ³³S. The new model uses input and feature sparsification, which reduces drastically the memory requirements and raises the prediction speeds, with similar prediction accuracy. Finally, it introduces error estimates in the shifts prediction, appending them in the extended text output.

19.07.2019: Version 1.1

• Addition of Sulfur containing structures in the training set. Changed to a sparse model to reduce memory requirements and calculation speed. Addition of error estimates to the predictions. Can now predict structures which contain H, C, N, O and S.

17.09.2018: Version 1.03 Beta

• Change the prediction procedure to reduce the memory requirements of the app.

27.06.2018: Version 1.02 Beta

• Implementation of a sparsified multi-scale-kernel, which reduces the calculation time and memory requirement while maintining similar accuracy.

02.05.2018: Version 1.0 Beta

• Implementation of the multi-scale-kernel version based on the description in Paruzzo, et al.. Can predict only structures containig H, C, N, and O.