Abstract

Using the example of a ML model for NMR chemical shieldings for molecular crystals from the CSD [1,2], various tricks of the trade are introduced. Including (i) the efficient estimation of uncertainties [3], (ii) sparsification of the chemical space [4], features, and similarity kernels [5] underlying KRR models, and (iii) the prediction of tensorial properties [6], these permit rendering comparatively simple KRR models (as outlined in previous talks in the MLDG series) practical and accurate for complex atomistic systems. The example of NMR chemical shieldings will also serve to touch upon some of the key limitations of our current ML approaches.

[1] Paruzzo et al., Nat Comm, 9, 4501 (2018) [2] Engel et al., PCCP , 21, 23385 (2019) [3] Musil et al., JCTC , 15, 906 (2019) [4] Willatt, Musil, and Ceriotti, PCCP , 20, 29661 (2018) [5] Rasmussen and Williams, Gaussian Processes for Machine Learning, MIT Press (2006) [6] Grisafi et al, PRL , 120, 036002 (2018)