The data contained inside ML_SJ.tar are organized in 3 folders and 3 compressed tar balls:


1) DECOMPOSITION_COEFFICIENTS.tar.gz:
   Compressed tar ball containing the decomposition coefficients for each molecule (train or test set) included in this work.
   The order of atomic orbitals corresponds to the one used in PySCF.
   The coefficients obtained with different density fitting metrics are labeled in the suffix of each file as follows:
     A) *.cj.dat   : J-metric
     B) *.cs.dat   : S-metric


2) ESP_GRID.tar.gz:
   Compressed tar ball containing the coordinates and weights of the grids points of isosurfaces (p0=1/16, 1/4, 1, 4, 32)
   used to compute electrostatic potential for each molecule (test set) included in this work.
     A) *.{0.0625,0.25,1,4,32}.5810.g.dat  : grid points coordinates (Bohr)
     B) *.{0.0625,0.25,1,4,32}.5810.w.dat  : grid points weights


3) GEOMETRIES:
     A) *.xyz       : the 2322 structures taken from the SSI subset of BFDb, that form the training and test sets, in the xyz format
     B) charges.dat : the list of molecules and their charges
     C) test.dat    : the list of molecules that form the test set along with their index numbers


4) PREDICTED_COEFFICIENTS.tar.gz:
   Compressed tar ball containing the predicted coefficients for each molecule (test set) included in this work.
   The coefficients obtained with different models (M0, MK), loss functions (JJ, JS, SJ, SS),
   and corrections for the number of electrons (refit, unit, scale) are labeled in the suffix of each file as follows:
     A) *.M0.{JJ,JS,SJ,SS}.dat                    : predictions of the M0 model without correction for the number of electrons
     B) *.M0.{JJ,JS,SJ,SS}.{refit,unit,scale}.dat : predictions of the M0 model corrected for the number of electrons
     C) *.MK.{JJ,SS}.{refit,unit,scale}.dat       : predictions of the MK model corrected for the number of electrons


5) REGRESSION_WEIGHTS:
   Contains the ML regression weights after training of each model (M0, MK) and loss function combination (JJ, JS, SJ, SS)
   for different training set fraction (1/8, 1/4, 1/2, 3/4, 1), along with the reference environments indices.
     A) M0.{JJ,JS,SJ,SS}.{0.125,0.25,0.5,0.75,1.0}.npy : M0 model
     B) MK.{JJ,SS}.{0.125,0.25,0.5,0.75,1.0}.npy       : MK model
     C) env.dat                                        : reference environments (sequentially numbered)


6) TABLES:
   Contains three OpenDocument sheets with errors in computed properties for each molecule (test set).
     A) dipole.ods : dipole moment errors (a.u.)
     B) dori.ods   : Walker--Mezey similarity indices
     C) esp.ods    : ESP errors (a.u.) on different surfaces