Bond breaking with auxiliary-field quantum Monte Carlo

Bond stretching mimics different levels of electron correlation and provides a challenging testbed for approximate many-body computational methods. Using the recently developed phaseless auxiliary-field quantum Monte Carlo (AF QMC) method, we examine bond stretching in the well-studied molecules BH and N$2$, and in the H${50}$ chain. To control the sign/phase problem, the phaseless AF QMC method constrains the paths in the auxiliary-field path integrals with an approximate phase condition that depends on a trial wave function. With single Slater determinants from unrestricted Hartree-Fock (UHF) as trial wave function, the phaseless AF QMC method generally gives better overall accuracy and a more uniform behavior than the coupled cluster CCSD(T) method in mapping the potential-energy curve. In both BH and N$_2$, we also study the use of multiple-determinant trial wave functions from multi-configuration self-consistent-field (MCSCF) calculations. The increase in computational cost versus the gain in statistical and systematic accuracy are examined. With such trial wave functions, excellent results are obtained across the entire region between equilibrium and the dissociation limit.

💡 Research Summary

This paper presents a comprehensive benchmark of the recently developed phaseless auxiliary‑field quantum Monte Carlo (AF QMC) method for describing bond stretching, a prototypical scenario where electron correlation changes dramatically and poses a severe challenge for approximate many‑body methods. The authors focus on three representative systems: the diatomic molecules BH and N₂, and a linear hydrogen chain of 50 atoms (H₅₀). For each system, they compare AF QMC results against essentially exact reference data—full configuration interaction (FCI) for the small molecules and density‑matrix renormalization group (DMRG) for the extended chain—as well as against the widely used coupled‑cluster singles‑doubles with perturbative triples, CCSD(T).

The AF QMC algorithm proceeds by projecting the ground state from a trial wave function |Ψ_T⟩ using a short‑time Trotter decomposition of the electronic Hamiltonian, followed by a Hubbard‑Stratonovich transformation that maps the two‑body interaction onto a set of auxiliary fields. The resulting imaginary‑time path integral is sampled by a branching random walk in Slater‑determinant space. The notorious sign/phase problem is controlled approximately by imposing a “phaseless” constraint that depends on the overlap with the trial wave function; the quality of |Ψ_T⟩ therefore directly determines the systematic bias. In the limit of an exact trial wave function the method becomes exact, and the only remaining error is statistical.

Two classes of trial wave functions are examined. First, a single‑determinant unrestricted Hartree–Fock (UHF) solution is used. Across all three systems, AF QMC with a UHF trial wave function yields potential‑energy curves that are at least as accurate as CCSD(T) near equilibrium and considerably more uniform as the bond is stretched. For BH, the UHF‑based AF QMC reproduces the FCI curve within ~3 mE_h up to five times the equilibrium bond length, whereas RCCSD(T) (based on a restricted Hartree–Fock reference) develops an unphysical dip beyond 2 R_e. For N₂, similar trends are observed: the UHF‑based AF QMC remains stable throughout the dissociation, while CCSD(T) suffers from divergence in the multi‑reference region. In the H₅₀ chain, a single‑determinant Hartree–Fock trial wave function already yields an energy profile in excellent agreement with DMRG, demonstrating the method’s scalability to large, low‑dimensional systems.

Second, the authors construct multi‑determinant trial wave functions from multi‑configuration self‑consistent‑field (MCSCF) calculations (specifically CASSCF with a (4e, 8o) active space for BH). Determinants are retained according to coefficient cut‑offs (e.g., 0.02, 0.01), resulting in trial wave functions containing from a few up to ~50 determinants. The inclusion of these determinants dramatically improves the phaseless constraint, reducing systematic errors to within ~1 mE_h of the exact reference for all bond lengths in BH. A quantitative efficiency metric η = (N_sample ε²)_MCSCF / (N_sample ε²)_UHF is introduced; values of η ranging from 0.04 to 0.08 indicate that, for the same statistical precision, the multi‑determinant trial wave function requires roughly an order of magnitude fewer Monte‑Carlo samples than the single‑determinant case. Similar improvements are reported for N₂, where the multi‑determinant AF QMC reproduces the FCI curve across the entire dissociation pathway.

The paper also discusses computational scaling. The cost of AF QMC grows linearly with the number of determinants in the trial wave function, but the dramatic reduction in statistical error often compensates for this overhead. For small molecules like BH, modest cut‑offs (e.g., retaining ~30 determinants) achieve near‑FCI accuracy with a modest increase in CPU time, while for larger systems the benefit of a multi‑determinant trial wave function must be balanced against the linear cost increase.

Overall, the authors draw several key conclusions: (1) Phaseless AF QMC with a simple UHF trial wave function already outperforms or matches CCSD(T) across a wide range of bond lengths, offering a more uniform description of both dynamic and static correlation. (2) Incorporating multi‑determinant trial wave functions derived from MCSCF dramatically reduces the systematic bias of the phaseless constraint, delivering near‑exact energies for bond‑stretching problems. (3) The method scales favorably with system size, as demonstrated by the accurate treatment of the 50‑atom hydrogen chain, and the statistical efficiency gains (η ≪ 1) make the approach practical even when a relatively large number of determinants is employed. (4) Because the only uncontrolled approximation is the quality of the trial wave function, AF QMC provides a systematic pathway to improve accuracy: better trial states (e.g., from selected CI, density‑matrix embedding, or tensor‑network methods) can be incorporated straightforwardly.

In summary, this work establishes phaseless AF QMC as a robust, highly accurate, and scalable alternative to traditional quantum‑chemical methods for challenging bond‑breaking problems, and it highlights the practical benefits of using multi‑determinant trial wave functions to capture static correlation while retaining the favorable computational scaling of the Monte‑Carlo framework.