Tolman Electronic Parameter Predictions from a Machine Learning Model Provide Insight into Phosphine Ligand Electronic Effects

Phosphines
are extremely important ligands in organometallic
chemistry,
and their donor or acceptor ability can be measured through the Tolman
electron parameter (TEP). Here, we describe the development of a TEP
machine learning model (called TEPid) that provides nearly instantaneous
calculation of experimentally calibrated CO vibrational stretch frequencies
for (R)3P–Ni0(CO)3 complexes.
This machine learning model with an error of less than 1 cm–1 (compared to density functional theory (DFT) calculated values)
was developed using >4,000 DFT calculated (R)3P–Ni0(CO)3 TEP values and 19 connectivity-based descriptors
associated with SMILES strings. We also built a web-based interface
to run the machine learning model where phosphine SMILES strings can
be entered and TEP values returned. We applied this TEPid model to
examine the donor and acceptor capabilities of phosphines in the large
Kraken phosphine database. This showed that the Kraken database is
skewed toward donor phosphines. In the same spirit of the Kraken database,
we generated tens of thousands of new experimentally based phosphines
that, when combined with Kraken phosphines, provide an electronically
balanced ligand library.