Abstract

As the workhorse of electronic structure methods, density functional theory (DFT) has become essential for molecule and materials discovery, deploying databases covering vast chemical and composition spaces, predicting chemical and electrochemical reactions, and developing machine learning potentials. The extensive application of DFT in chemistry and materials science requires predictions achieving “chemical accuracy,” a goal hindered by the unknown functional form for the exact exchange and correlation (XC) energy. A recent meta-generalized gradient approximation, the restored-regularized strongly constrained and appropriately normed (r²SCAN) XC functional, fulfils 17 exact constraints of the XC energy, and has significantly boosted the accuracy of predictions for molecules and materials. However, r²SCAN still appears inadequate at predicting material properties of open d and f transition-metal strongly correlated compounds, including band gaps, magnetic moments, and oxidation energies. Prediction inaccuracies of r²SCAN arise from functional and density-driven errors, mainly resulting from the DFT self-interaction error. Here, we propose a novel method termed r²SCANY@r²SCANX to mitigate the pernicious self-interaction error of XC functionals for the accurate simulations of electronic, magnetic, and thermochemical properties of transition metal oxides. r²SCANY@r²SCANX utilizes different fractions of exact Hartree-Fock exchange: X to define the electronic density, and Y to set the density functional approximation for the energy. We show that r²SCANY@r²SCANX simultaneously addresses functional-driven and density-driven inaccuracies, mitigating the self-interaction error in DFT. Building just on 1 (or maximum 2) universal parameters, we demonstrate that r²SCANY@r²SCANX improves upon the r²SCAN predictions for 18 highly correlated oxides and even outperforms the highly parameterized DFT(r²SCAN)+U method –the state-of-the-art approach to predict strongly correlated materials. We demonstrate that the typical O₂ overbinding inaccuracies of the local density and the gradient XC approximations, with errors ranging between –2.2 and –1.0 eV/O₂, reduce to –0.3 eV/O₂ for r²SCAN, and further reduce to less than∼0.03 eV/O₂ using any combination of X in r²SCAN10@r²SCANX, including X=0. Prediction uncertainties for oxidation energies and magnetic moments of transition metal oxides are significantly reduced by r²SCAN10@r²SCAN50 and band gaps with r²SCAN10@r²SCAN. r²SCAN10@r²SCAN50 diminishes the density-driven error present in r²SCAN and r²SCAN10. We demonstrate that the computationally efficient r²SCAN10@r²SCAN is nearly as accurate as the global hybrid r²SCAN10 for oxidation energies. This suggests that accurate energy differences can be achieved through rate-limiting self-consistent iterations and geometry optimizations using the efficient r²SCAN. Subsequently, a more expensive nonlocal functional, such as a hybrid or self-interaction correction, can be applied in a fast-to-execute single post-self-consistent calculation, as in r²SCAN10@r²SCAN. Compared to computationally demanding r²SCAN10, r²SCAN10@r²SCAN appears to be 12 to 165 times faster for the iron oxides in our test set.