We employed our recently proposed multireference approach CASPT2/CC to calculate the quintet-triplet gaps E TQ of a series of non-heme Fe IV =O species and subsequently used these results to benchmark density functional theory (DFT) as well as two variants of local coupled-cluster approaches (DLPNO-CCSD(T) and LUCCSD(T0)).We showed that current implementations of the local coupled-cluster method are not su ciently accurate. DLPNO-CCSD(T) systematically overstabilizes the quintet state, whereas LUCCSD(T0) overestimates the triplet one. This sort of systematic bias may be helpful in improving local correlation methods, and can also be used as the basis for a simple correction scheme.
1The importance of non-heme iron enzymes has been realized in the past years, 1-8 leading to the synthesis of numerous complexes 9-11 mimicking the reactions performed by these enzymes, as well as their study by means of both experiment and theory. 1,[12][13][14] On the theoretical side, most of these studies have employed density functional theory (DFT), usually with the B3LYP functional, 8,[15][16][17][18][19] in order to gain insights into these complexes and their multiple spin states. However, it is known that DFT can fail to properly describe the relative energy between di↵erent spin states of transition metal (TM) complexes. [20][21][22][23] A possible solution for this problem might come from the use of more accurate ab initio methods, but these have been much less used due to their high computational costs. For instance, the "gold-standard" canonical CCSD(T) can be used only on small model systems, [24][25][26] while conventional multireference approaches such as CASSCF/CASPT2 27 or CASSCF/NEVPT2 28 are normally applicable for small to medium-sized mononuclear TM complexes. This has stimulated the development of novel approaches which significantly improve computational e ciency without sacrificing accuracy, such as the local coupled-cluster (CC) methods (DLPNO-CCSD(T) [29][30][31][32][33][34] and LUCCSD(T) [35][36][37][38][39] ) or DMRG-based methods, [40][41][42][43][44][45][46][47][48][49][50][51][52] or Stochastic-CASSCF approach. 53 However, in some recent studies, 26,54,55 it has been demonstrated that both CASPT2 and local CC methods might produce results of insu cient accuracy for TM complexes. The errors of the former are attributed to an erroneous description of the metal semicore (3s,3p) correlation e↵ect in the PT2 treatment; 54 or (in some di cult cases) to deficiencies of PT2 to describe higher-order excitations 56 and of the active space to describe the charge-transfer configuration between the metal and ligands. 55 In a recent work, 57 we proposed a composite approach, referred to as CASPT2/CC, in which the (3s,3p) correlation e↵ect is separately treated with ROHF-CCSD(T), using a very small basis set except on the metal centre. Valence correlation was calculated with CASSCF/CASPT2, in combination with an extensive correlation consistent (cc-) basis set. It has been demonstrated that CASPT2/CC is e cient and ca...