Abstract: The multiconfigurational self-consistent field theory is considered the standard starting point for almost all multireference approaches required for strongly correlated molecular problems. The limitation of the approach is generally given by the number of strongly correlated orbitals in the molecule, since its cost will grow exponentially with this number. We present a new multiconfigurational self-consistent field approach, wherein linear determinant coefficients of a multiconfigurational wave function are optimized via the stochastic full configuration interaction quantum Monte Carlo technique at greatly reduced computational cost, with nonlinear orbital rotation parameters updated variationally based on this sampled wave function. This extends this approach to strongly correlated systems with far larger active spaces than it is possible to treat via conventional means. By comparison with this traditional approach, we demonstrate that the introduction of stochastic noise in both the determinant amplitudes and the gradient and Hessian of the orbital rotations does not preclude robust and reliable convergence of the orbital optimization. It can even improve the ability to avoid convergence to local minima in the orbital space, and therefore aid in finding variationally lower-energy solutions. We consider the effect on the convergence of the orbitals as the number of walkers and the sampling time within the active space increases, as well as the effect on the final energy and error. The scope of the new protocol is demonstrated with a study of the increasingly strongly correlated electronic structure in a series of polycyclic aromatic hydrocarbons, up to the large coronene molecule in a complete active space of 24 π electrons in 24 orbitals, requiring only modest computational resources.
Introduction
In quantum chemistry, most "gold standard" methods, such as CCSD(T), rely on a single-particle picture where electrons move in an average field of others. While effective for many molecules, this approximation fails for "strongly correlated" systems, including transition metals, bond-breaking scenarios, and large aromatic molecules. To address these, researchers use the Complete Active Space Self-Consistent Field (CASSCF) method, which considers all possible electron configurations within a specific set of "active" orbitals.
However, CASSCF faces an "exponential wall": the number of configurations grows factorially with the number of orbitals. Traditionally, this has limited active spaces to approximately 16 orbitals, beyond which memory requirements (reaching hundreds of terabytes) and computational time become prohibitive. This paper introduces FCIQMC-CASSCF, a stochastic approach that breaks through this limitation by combining Monte Carlo sampling with traditional orbital optimization.
Methodology
The proposed method utilizes a two-step "macroiterative" approach to optimize the molecular wave function:
- Stochastic Sampling (FCIQMC): Instead of solving the massive configuration matrix exactly, the method employs Full Configuration Interaction Quantum Monte Carlo. This technique uses a population of "signed walkers" that spawn, die, and annihilate across the space of possible electron arrangements. This allows for the sampling of the wave function and the computation of Reduced Density Matrices (RDMs) without the need to store the full, gargantuan configuration vector.
- Deterministic Orbital Optimization: The sampled RDMs are used to calculate the energy gradient and Hessian. A standard variational update is then performed to rotate and optimize the molecular orbitals.
A key finding of this work is that the inherent stochastic noise from the Monte Carlo sampling does not hinder convergence. Instead, it acts as a "shaking" mechanism that helps the algorithm escape local energy minima to find the true global minimum.
Validation and Benchmarking
The researchers validated the FCIQMC-CASSCF method against exact deterministic results for small molecules like benzene and naphthalene. As shown in Figure 1, the stochastic approach matches deterministic energy convergence with sub-millihartree precision.
Beyond simple benchmarks, the method was applied to study the torsional barrier of ethene—a classic test for strong correlation—where it successfully smoothed out the unphysical "energy cusps" that plague simpler theories like Hartree-Fock or MP2.
Application to Large Systems: The Coronene Study
The true power of the method was demonstrated through a study of polycyclic aromatic hydrocarbons (PAHs), culminating in the analysis of coronene. This required a CAS(24,24) active space (24 electrons in 24 orbitals), containing roughly $10^{12}$ determinants. A traditional CASSCF approach would require approximately 100 TB of RAM; the FCIQMC-CASSCF method achieved convergence using only 3 GB of distributed memory.
Figure 2 illustrates the convergence for coronene with varying walker populations ($N_w$). Interestingly, calculations with fewer walkers (and thus more noise) were sometimes able to find lower-energy orbital solutions by jumping out of meta-stable local minima, highlighting "noise as a feature" of the stochastic optimization process.
Conclusion and Future Work
FCIQMC-CASSCF enables the study of strongly correlated systems with active spaces far larger than previously possible. Beyond memory efficiency and scalability, the method allowed for physical discoveries, such as identifying that reactive electron density in coronene accumulates primarily at the outer carbon "edges."
While the method still exhibits weak exponential scaling, it significantly pushes the boundaries of multireference quantum chemistry. Future work will focus on extending the protocol to "state-averaged" CASSCF for photochemistry and exploring simultaneous optimization of coefficients and orbitals to further enhance efficiency.
Summary updated