Assessing the Metal–Metal Interactions in a Series of Heterobimetallic Nb/M Complexes (M = Fe, Co, Ni, Cu) and Their Effect on Multielectron Redox Properties

Abstract

A one-pot synthetic procedure for a series of bimetallic Nb/M complexes, Cl−Nb( i PrNPPh 2 ) 3 M−X (M = Fe (2), Ni (4), Cu (5)), is described. A similar procedure aimed at synthesizing a Nb/Co analogue instead affords i PrNNb( i PrNPPh 2 ) 2 (μ-PPh 2 )Co−I (3) through cleavage of one phosphinoamide P−N bond under reducing conditions. Complexes 4 and 5 are found to have short Nb-M distances, corresponding to unusual metal−metal bonds between Nb and these first row transition metals. For comparison, a series of heterobimetallic ONb( i PrNPPh 2 ) 3 M−X complexes (M = Fe (7), Co (8), Ni (9), Cu ( 10)) was synthesized. In these complexes, the Nb V center is engaged in sufficient π-bonding to the terminal oxo ligand to remove the driving force for direct metal−metal interactions. A comparison of the cyclic voltammograms of 2 and 4−10 reveals that the presence of a second metal shifts the redox potentials of both Nb and the late metal center anodically, even when direct metal−metal interactions are not present.

Deciphering Cryptic Behavior in Bimetallic Transition-Metal Complexes with Machine Learning

“…For the linear model with mm-RAC coefficients over the smaller oxidation state set, the covalent radius of the metal and its nearest neighbors were key along with the electronegativity of the metal's nearest neighbors (Figure 5). This leads the model to predict oxidation potentials for set-aside test L 4 M2M1 complexes 91,92 to be of the order of Fe ~ Co < Cu for M1 in either M2 = Cr or Nb cases because across this series both covalent radius decreases and electronegativity increases, in qualitative accordance with chemical trends (Supporting Information Table S5). The model also correctly predicts that the difference between Fe and Cu M1 complex oxidation potentials will be larger in the case where M2 = Nb than Cr in part because Nb is significantly larger (i.e., in the S1 contribution to the linear model) despite having a comparable electronegativity to Cr (Figure 1 and Supporting Information Table S1).…”

“…For most other cases, training errors are significantly lower (typically less than 0.2-0.3 V), and model prediction accuracy is good (Figure 2). To confirm the generality of our approach, we test on set-aside literature results 91,92 for seven additional complexes with a distinct ligand structure, L 4 , with N,P-coordination that was not in the training data and include previously unseen (i.e., Nb) elements in the M2 site (Figure 1 and Supporting Information Table S5). Across this set of seven new complexes, the MAE of our model is only slightly worse at around 0.38 V (Figure 2 and Supporting Information Table S3).…”

“…100 This final oxidation potential set consisted of 35 complexes, 22−25,37,82−92 seven of which were held out as set aside test data. 91,92 Both multiple linear regression (MLR) and kernel ridge regression (KRR) models were trained in scikit-learn. 101 The (e)mm-RAC features were applied to the CSD structures using the molSimplify 54,102 code.…”

“…All 35 oxidation potential set molecules are included in this set, which was further refined by searching for abstracts associated with structures that contained cyclic voltammetry keywords using the pybliometrics code . This final oxidation potential set consisted of 35 complexes, − ,,− seven of which were held out as set aside test data. , …”

“…To confirm the generality of our approach, we test on set-aside literature results , for seven additional complexes with a distinct ligand structure, L 4 , with N,P-coordination that was not in the training data and include previously unseen (i.e., Nb) elements in the M2 site (Figure and Table S5). Across this set of seven new complexes, the MAE of our model is only slightly worse at around 0.38 V (Figure and Tables S4 and S5).…”