Reverse Quantum Annealing for Ground-State Energy Calculation in Molecule

AI summary

Overview

Research area: Quantum computing for chemistry — specifically quantum annealing (QA) applied to relativistic molecular electronic-structure calculations.

Technical level: Intermediate. The paper assumes familiarity with second quantization, QUBO formulations, and quantum annealing hardware, but its central result (a refinement step that lowers errors on a quantum annealer) is easy to grasp.

Scope (one sentence): The paper tests whether reverse quantum annealing (RQA), run over twelve different schedules, can refine forward-annealing solutions from the Quantum Annealing Eigensolver (QAE) for five alkaline-earth monofluoride radicals in two active spaces and push the resulting energies toward exact Full Configuration Interaction (FCI) references.

What This Paper Is About

Quantum annealers are heuristic optimizers: when a molecular Hamiltonian is encoded as a binary optimization problem, the lowest-energy sample returned by a forward anneal is not guaranteed to be the true ground state, and errors grow as the encoded problem gets larger. This paper asks whether a second, cheaper pass — reverse annealing that starts from the forward-annealing answer, briefly re-enters a more quantum regime, and returns to the problem Hamiltonian — can systematically recover the missing accuracy. The goal is to quantify how much of the residual error relative to exact FCI diagonalization such a refinement can remove for BeF, MgF, CaF, SrF, and BaF, and which schedule parameters work best.

Key Contributions

1. RQA as a refinement layer for relativistic QAE. The author applies reverse quantum annealing on top of a forward-annealing QAE workflow built from the Dirac–Coulomb Hamiltonian, for five open-shell alkaline-earth monofluoride radicals (BeF, MgF, CaF, SrF, BaF) in (3e,8o) and (7e,14o) active spaces.

2. A twelve-schedule design study. Twelve reverse-annealing schedules (S1–S5 pause-type, S6–S12 V-shaped) are compared, spanning reversal points from s′ = 0.35 to s′ = 0.65 with different dwell times, so reversal depth and schedule shape can be assessed while holding the molecular Hamiltonian and active space fixed.

3. Exact FCI benchmarking with a defined error-reduction metric. Forward- and reverse-annealing energies are compared against exact FCI diagonalization in the identical relativistic spinor determinant basis, and a percentage residual-error reduction η is defined and reported for all ten molecule/active-space combinations.

4. A concrete accuracy result and a stated limitation. RQA eliminates 82.3% to 98.1% of residual forward-annealing errors, reaching sub-milliHartree errors for BeF in (3e,8o) (0.305 mHa) and SrF in (7e,14o) (0.596 mHa), while the recoverable relative improvement shrinks by roughly a factor of 4.5 when the active space is enlarged.

Main Findings

• Best schedule is S11 (s′ = 0.40) in both active spaces. Mean signed relative energy change is −0.690% in (3e,8o) and −0.151% in (7e,14o); S9 (s′ = 0.45) is second in both, at −0.617% and −0.129%.

• An intermediate-reversal "Goldilocks" window. The strongest schedules cluster around s′ = 0.35–0.45, where transverse-field fluctuations are strong enough to escape a local basin but the forward-annealing information is not fully erased. S8 (s′ = 0.65) reproduces the forward-annealing energy exactly in both active spaces, and S5 and S12 also show zero mean change in (7e,14o).

• Larger active space, smaller recoverable gain. Averaged over all molecule–schedule observations, ΔE is −0.409% for (3e,8o) versus −0.071% for (7e,14o); at the best schedule the values are −0.690% and −0.151%, a decrease by a factor of approximately 4.5.

• Strong molecule dependence. In (3e,8o), molecule-averaged lowering is MgF −0.568%, BaF −0.475%, CaF −0.384%, SrF −0.356%, BeF −0.262%. In (7e,14o) the order changes: MgF −0.138%, BeF −0.098%, CaF −0.061%, BaF −0.036%, SrF −0.022%.

• The best schedule per molecule is not universal. In (3e,8o) the lowest observed changes come from S1 for BaF (−1.054%) and BeF (−0.554%), S4 for CaF (−0.759%), and S11 for MgF (−1.292%) and SrF (−0.712%). In (7e,14o) they come from S2 for BaF (−0.085%), S9 for BeF (−0.225%), and S11 for CaF (−0.139%), MgF (−0.335%), and SrF (−0.051%).

• RQA usually helps but is stochastic. Across the 60 observations per active space, (3e,8o) has 50 improved, 8 unchanged, and 2 higher-energy cases (BeF/S11 at +0.043% and MgF/S4 at +0.218%); (7e,14o) has 45 improved, 15 unchanged, and 0 higher. Standard deviation for S11 across molecules is approximately 0.505% in (3e,8o) and 0.115% in (7e,14o).

• Exact FCI benchmark, (3e,8o). BeF: E_FCI −2.879126 Ha, E_FA −2.862953 Ha, E_RA −2.878821 Ha (S1), δ_FA 16.173 mHa → δ_RA 0.305 mHa, η = 98.12%. MgF: −2.548084 / −2.510064 / −2.542488 Ha (S11), 38.021 → 5.596 mHa, η = 85.28%. CaF: −2.372715 / −2.351043 / −2.368891 Ha (S4), 21.672 → 3.825 mHa, η = 82.35%. SrF: −2.318931 / −2.299062 / −2.315439 Ha (S11), 19.869 → 3.492 mHa, η = 82.42%. BaF: −2.221598 / −2.194220 / −2.217353 Ha (S1), 27.378 → 4.246 mHa, η = 84.49%.

• Exact FCI benchmark, (7e,14o). BeF: −17.081188 / −17.035968 / −17.074222 Ha (S9), 45.220 → 6.966 mHa, η = 84.59%. MgF: −16.011878 / −15.952295 / −16.005813 Ha (S11), 59.583 → 6.065 mHa, η = 89.82%. CaF: −15.512783 / −15.488854 / −15.510336 Ha (S11), 23.929 → 2.447 mHa, η = 89.78%. SrF: −15.254824 / −15.246472 / −15.254229 Ha (S11), 8.352 → 0.596 mHa, η = 92.87%. BaF: −14.740778 / −14.726527 / −14.739000 Ha (S2), 14.251 → 1.778 mHa, η = 87.53%.

• Where chemical accuracy is reached. The paper uses the standard threshold of 1.0 kcal/mol ≈ 1.594 mHa. Forward annealing leaves every system between 8.35 mHa and 59.58 mHa of FCI; RQA brings BeF (3e,8o) to 0.305 mHa and SrF (7e,14o) to 0.596 mHa, inside the sub-milliHartree regime, and BaF (7e,14o) at 1.778 mHa to the threshold. The paper describes the (7e,14o) mean residual-error reduction as approximately 88%, and its conclusion describes that active space as eliminating approximately 85–93% of residual error.

Methodology in Plain English

The workflow has eight stages. First, a relativistic Dirac–Coulomb Hamiltonian is built for each molecule from one- and two-electron integrals in a relativistic spinor basis. Second, the eigenvalue problem is converted into an optimization problem by expanding the wave function in determinants and adding a Lagrange multiplier λ for normalization, giving a functional minimized over the coefficients. Third, each coefficient in [−1, 1] is encoded in K = 10 binary variables with a fixed-point scheme, giving a discretization resolution of Δc = 2⁻⁹ ≈ 1.95 × 10⁻³ and a QUBO with N × K variables. Fourth, because the hardware cannot embed one large dense QUBO, the problem is split into smaller sub-QUBOs, with coefficients ordered by a first-order perturbation theory estimate C_α ~ H_0α/(H_00 − H_αα). Fifth, each sub-QUBO is minor-embedded on the D-Wave Advantage system 4.1 Pegasus graph and forward-annealed with 1000 reads at 20 μs. Sixth, the lowest-energy sub-QUBO samples are concatenated into one starting configuration. Seventh, that configuration seeds reverse annealing — the anneal parameter goes from the problem endpoint to a reversal point s′ and back to 1 — under each of the twelve schedules, with 1000 reads each. Eighth, the refined bitstrings are decoded back into real coefficients, the couplings lost in partitioning are restored, and the variational energy is evaluated by scanning λ over an interval slightly below and above the Dirac–Fock energy.

The classical reference energies come from Dirac–Fock and relativistic coupled-cluster singles and doubles calculations in DIRAC22, with exact references from constructing and fully diagonalizing the FCI Hamiltonian in the Slater–Condon determinant basis. Basis sets were uncontracted Gaussian-type orbitals with kinetic balance: correlation-consistent polarized valence sets from the EMSL Basis Set Exchange for Be, Mg, Ca, and F, and dyall.c4v with additional tight primitives for Sr and Ba.

Why This Matters

Impact on research. The paper provides a quantitative, FCI-anchored demonstration that the bottleneck in annealer-based molecular eigensolvers is not only the encoding but the sampling stage, and that a cheap second pass can recover most of the lost accuracy. It also documents a clear scaling warning: the same refinement machinery recovers far less relative improvement at (7e,14o) than at (3e,8o), and the best schedule is molecule dependent, which argues for small schedule ensembles rather than one universal schedule.

Real-world applications (as framed by the paper):

• Relativistic molecular electronic-structure calculations, where accurate ground-state energies underpin bonding, stability, and molecular property predictions.
• Molecular vibrational spectra and energies of molecular electronic states, cited as broader QAE use cases.
• Permanent electric dipole moment computations and relativistic fine-structure splitting in atomic ions, which are precision tests of fundamental physics.
• Precision-measurement molecules: SrF, described as the first molecule to be laser-cooled, and BaF, described as a promising candidate for precision measurements in fundamental physics.

Industry relevance. The work runs entirely on commercially available annealer hardware (D-Wave Advantage system 4.1) using the D-Wave Ocean software stack, so it speaks directly to the practical question of what near-term annealing hardware can deliver for quantum chemistry and, more generally, for binary optimization problems that can be mapped to QUBOs.

Future Directions

• Scaling the active space. The recoverable relative improvement drops from −0.409% to −0.071% on going from (3e,8o) to (7e,14o); the paper attributes part of this to embedding and stochastic errors in the larger binary problem, leaving open how far the approach extends.
• Schedule ensemble strategies. Because the best schedule differs by molecule (S1, S4, S9, S11, and S2 all appear as winners), a principled way to choose or combine schedules per problem is an open practical question.
• Reducing embedding and chain overhead. The paper frames intermediate reversals as a way around convergence bottlenecks arising from embedding and chain effects; improving the embedding itself, or the sub-QUBO partitioning that produces the initial configuration, is a natural next step.
• Moving beyond energies. The introduction lists dipole moments, fine-structure splittings, and vibrational spectra as QAE targets, but the reported benchmarks cover only ground-state energies, so extending the RQA refinement to these properties remains untested.

Target Audience

Researchers and graduate students in quantum computing for chemistry, relativistic quantum chemistry, and quantum annealing hardware applications. It is most useful to readers already comfortable with second quantization and QUBO encodings who want a concrete, benchmarked picture of how much a reverse-annealing refinement pass buys on real hardware — and where that benefit thins out as problem size grows. Readers interested primarily in gate-model variational quantum eigensolvers will find the hardware paradigm different, but the FCI benchmarking methodology and the schedule-engineering lessons transfer.

Authors’ abstract

We present the results of our reverse quantum annealing computations of the ground-state energies of alkaline-earth monofluoride molecules. By partitioning the Dirac--Coulomb electronic Hamiltonian into sub-QUBOs and refining forward-annealing solutions through optimized reverse-annealing schedules, residual configuration errors are systematically minimized. The electronic ground-state wave functions and energies are obtained using the Quantum Annealing Eigensolver algorithm. This work provides a pathway for achieving sub-milliHartree chemical accuracy and advancing molecular electronic structure calculations in the quantum annealing paradigm.