For evaluation purposes, the following datasets are employed:

The first dataset is based on synthetic diffusion-weighted data for 10, 20, 40 and 80 gradient directions and b = 3000 s/mm^{2}. It is simulated for different Rician noise levels (SNR = ^ = {10, oo}), which represent MRI scans with a high noise level and no noise, respectively. The signal is simulated based on the Multi-Tensor Model [15]. Tensor eigenvalues are set according to real DW-MRI data [3], while the number of compartments is chosen randomly between 1 and 3. Corresponding volume fractions range from 0.2 to 1 and sum up to 1. The training set consists of

40.000 voxels and the test set of 10,000 voxels for which the signal was simulated for each gradient set and noise level.

The second dataset contains data from 100 different uncorrelated healthy subjects from the Human Connectome Project (HCP). From each subject, 5000 white matter voxels were extracted randomly, resulting in 500,000 voxels in total. All three shells (b = {1000,2000, 3000} s/mm^{2}), with 90 gradient directions each, are used in this work, being either training data or target data for prediction. The dataset is split into a training set containing 450,000 voxels from 90 subjects and a test set consisting of

50.000 voxels from 10 subjects, which ensures that each subject is included in either the training or the test set. In order to create a subsampled dataset containing less, but still equidistantly distributed gradient directions, SH are fitted to the signal and subsampled by a new gradient set with 15 gradient directions. 15 gradient directions are chosen because it is the minimum number of measurements that are required to fit SH of order 4, if coefficients are calculated through matrix inversion [6].