A well-performing machine learning (ML) model is obtained
by using
proper descriptors and artificial neural network (ANN) algorithms,
which can quickly and accurately predict activation free energy in
hydrogen atom transfer (HAT)-based sp
3
C–H activation.
Density functional theory calculations (UωB97X-D) are used to
establish the reaction system data sets of methoxyl (CH
3
O·), trifluoroethoxyl (CF
3
CH
2
O·),
tert
-butoxyl (tBuO·), and cumyloxyl (CumO·) radicals.
The simplified Roberts’ equation proposed in our recent study
works here [
R
2
= 0.84, mean absolute error
(MAE) = 0.85 kcal/mol]. Its performance is comparable with univariate
Mulliken-type electronegativity (χ) with the ANN model. The
ANN model with bond dissociation free energy, χ, α-unsaturation,
and Nolan buried volume (%
V
buried
) successively
improves
R
2
and MAE to 0.93 and 0.54 kcal/mol,
respectively. It reproduces the test sets of trichloroethoxyl (CCl
3
CH
2
O·) with
R
2
= 0.87 and MAE = 0.89 kcal/mol and accurately predicts the relative
experimental barrier of the HAT reactions with CumO· and the
site selectivity of CH
3
O·.