Abstract
One of the key challenges of quantum-chemical multi-configuration methods is the necessity to manually select orbitals for the active space. This selection requires both expertise and experience and can therefore impose severe limitations on the applicability of this most general class of ab initio methods. A poor choice of the active orbital space may yield even qualitatively wrong results. This is obviously a severe problem, especially for wave function methods that are designed to be systematically improvable. Here, we show how the iterative nature of the density matrix renormalization group combined with its capability to include up to about 100 orbitals in the active space can be exploited for a systematic assessment and selection of active orbitals. These benefits allow us to implement an automated approach for active orbital space selection, which can turn multi-configuration models into black box approaches.
Introduction
Multi-configuration methods are essential for describing molecules with complex electronic structures, such as transition metal complexes or molecules undergoing bond dissociation. However, these methods traditionally require the user to manually select a subset of "active" orbitals where the most significant electron correlations occur. This selection process, often described as a "dark art," relies heavily on researcher intuition and trial-and-error. A suboptimal choice can lead to qualitatively incorrect results, undermining the systematic improvability of wave function methods. To address this, we present an automated protocol that leverages the Density Matrix Renormalization Group (DMRG) and entanglement-based measures to make active space selection a systematic, "black-box" procedure.
Entanglement-Based Selection Methodology
The core of the automated protocol utilizes the DMRG algorithm's ability to handle large active spaces (up to 100 orbitals). The selection is guided by quantum entanglement measures: the single-orbital von Neumann entropy ($s_i(1)$) and mutual information ($I_{ij}$). These metrics quantify the degree of interaction between orbitals.
A significant finding of this work is that fully converged wave functions are not required to identify the optimal active space. As illustrated in Figure 1, "cheap" exploratory DMRG calculations with low bond dimensions ($m$) and few sweeps provide entanglement patterns that are qualitatively identical to those from expensive, fully converged calculations. The protocol identifies "plateaus" in plots of orbital count versus entropy thresholds, providing a mathematical criterion for defining the active space and removing human bias.
Case Studies and Results
The protocol was evaluated on several challenging molecular systems. For the permanganate ion ($ ext{MnO}_4^-$), the method automatically identified the 17-orbital active space previously determined through extensive manual study.
In the study of the Mn(salen) complex (Figure 2), the protocol produced a unified 21-orbital active space for multiple spin states. This resulted in a singlet-triplet splitting of 0.6 kcal/mol, which is in excellent agreement with high-level benchmarks and corrects the large errors (40-45 kcal/mol) seen in restricted CASSCF models. Additionally, for the $ ext{Cu}_2 ext{O}_2^{2+}$ isomerization, the method consistently selected a 28-orbital space along the reaction coordinate, yielding relative energies much closer to gold-standard coupled-cluster results than traditional small-space methods.
Conclusion
By automating the selection of active orbital spaces through entanglement measures, this approach transforms multi-configuration models into more accessible and reliable tools. While the method currently requires an initial large guess of orbitals, it significantly reduces the expertise required to perform high-accuracy quantum chemical calculations and provides a robust foundation for including dynamic correlation.
Summary updated