Machine-Learned Compact Subspace Generation for Quantum Selected Configuration Interaction within Density Matrix Embedding Framework

AI summary

Overview

Research area: Quantum chemistry on near-term quantum computers — specifically hybrid quantum-classical electronic structure methods that combine quantum sampling with classical diagonalization, machine learning (Restricted Boltzmann Machines), and density matrix embedding (DMET) to treat a protein–ligand complex.

Technical level: Advanced. The paper assumes familiarity with second quantization, selected configuration interaction, DMET embedding, and generative neural networks; the abstract and conclusions are accessible to an intermediate reader.

Scope: The paper introduces QSCI-RBM — a Restricted Boltzmann Machine that generates compact determinant subspaces from hardware-sampled bitstrings — validates it on two small molecules (ethylene, methanolamine), and embeds it in DMET to compute the energy of the Carmofur / SARS-CoV-2 main protease complex split into 11 fragments.

What This Paper Is About

Quantum Selected Configuration Interaction (QSCI), also called Sample-based Quantum Diagonalization (SQD), works by sampling bitstrings from a quantum circuit and then exactly diagonalizing the Hamiltonian in the subspace of sampled determinants. Its accuracy depends entirely on whether the sampled subspace contains the chemically important configurations, and existing "configuration recovery" schemes only enforce symmetry (electron number, spin) without ensuring that the most physically relevant determinants are kept — so subspaces become unnecessarily large and classical diagonalization becomes expensive. This paper replaces that heuristic recovery with a machine-learned generator: a Restricted Boltzmann Machine trained on the quantum samples proposes high-probability determinants that are added back into the subspace iteratively. The authors then plug this solver into Density Matrix Embedding Theory and apply it to a drug-bound protein pocket.

Key Contributions

1. An RBM-guided iterative subspace expansion protocol (QSCI-RBM) in which a Restricted Boltzmann Machine is trained on symmetry-filtered hardware bitstrings and used for Gibbs sampling to propose new, symmetry-valid, not-yet-present determinants at every iteration, with a persistent "determinant memory" updated by a CI-coefficient threshold (top-K above τ_sv).

2. A spin-string proliferation step used deliberately as a second discovery route: rather than diagonalizing only over explicitly sampled determinants, the unique α- and β-string sets are pooled and their full Cartesian product is formed, surfacing cross-paired configurations that were never sampled by hardware or generated by the RBM.

3. A design choice to train the RBM with equal weighting over all determinants in the memory (excluding the Hartree–Fock reference), so the model functions as an unbiased sample generator over a pre-filtered dominant sector rather than as a fit to the exact ground-state probability distribution |c_φ|².

4. Integration of QSCI-RBM as the high-level impurity solver inside the DMET self-consistency loop (DMET-QSCI-RBM), where the global chemical potential is updated by a Newton–Secant procedure, and a demonstration on a pharmaceutically relevant system fragmented into 11 impurities.

Main Findings

• Ethylene (C₂H₄, STO-3G, CAS = 9,018,009 determinants, 100K shots, IBM Heron Fez): The absolute energy error fell from roughly 200 mHa at iteration 1 to 1.11 mHa at iteration 28, below the 1.6 mHa chemical-accuracy threshold and below the CCSD reference error of 1.22 mHa. The error later settled to roughly 0.4 mHa.

• Compactness on ethylene: Chemical accuracy was first achieved while accessing only 2.12% of the full symmetry space; the accessed fraction grew from roughly 0.5% and plateaued with mild oscillations around roughly 2.4% after iterations 28–30.

• Methanolamine (CH₅NO, STO-3G, CAS (26e, 20o), Hilbert space ≈ 6.01 × 10⁹, roughly 670× larger than ethylene): The error dropped from roughly 200 mHa to 1.47 mHa at iteration 29, where chemical accuracy was first crossed; the authors note the CCSD reference error for this system falls just inside the chemical-accuracy region. The figure caption quotes the chemical-accuracy band as 1.59 mHa.

• Compactness grows with system size: On CH₅NO, chemical accuracy was reached while accessing only 0.06% of the symmetry space (versus 2.12% for ethylene), with the subspace reaching roughly 0.085% by iteration 40 and no clear saturation — a more favourable compactness–accuracy trade-off as the Hilbert space grows.

• RBM recovers correlation beyond CCSD (ethylene): The converged QSCI-RBM error of 1.11 mHa is below the CCSD error of 1.22 mHa, meaning the hardware-sampled, RBM-guided subspace captured more correlation energy than a classical coupled-cluster singles and doubles treatment.

• Protein–ligand result (abstract): DMET-QSCI-RBM reached energies within the chemical-accuracy threshold while accessing only approximately 4% of the configuration subspace, whereas standard DMET-SQD failed to reach chemical accuracy while accessing up to 20% of the subspace, even though the chemical potential itself nearly converged.

• Test systems and resources for the protein work: The Carmofur/Mpro binding pocket (residues within 3 Å of the ligand, defined via MFCC capping) was split into 11 fragments of 2–7 atoms (mostly 5–6), described in 6-31G with impurities of 44–152 spin orbitals; each was reduced to a uniform (16e, 16o) active space = 32 qubits under Jordan–Wigner, giving a per-fragment symmetry space of 165,636,900 determinants.

• Hardware-sampling characterisation: The authors report that LUCJ-prepared samples populate excitation orders roughly four through six relative to the Hartree–Fock reference with substantial shot density — a range inaccessible to a doubles-restricted perturbative seed such as MP2 — although they explicitly state it is not established whether these orders contribute meaningfully to the converged wavefunction.

• Caveats stated by the authors: RBM hyperparameters (learning rate, hidden units, Gibbs chain length, singular-value threshold) were chosen empirically with no systematic hyperparameter optimisation. The DMET-QSCI-RBM results are reported at a halted μ-self-consistency point, and DMET-SQD baselines used two S-CoRe recovery iterations per impurity.

Methodology in Plain English

The pipeline has three nested layers.

Innermost — the quantum sampler. A single LUCJ (Local Unitary Cluster Jastrow) circuit — orbital rotations plus spin-resolved Jastrow correlators acting on a Hartree–Fock reference — is run on hardware, and the α and β registers are measured to produce bitstrings. These are immediately filtered for correct electron number (N_α + N_β = N_elec) and zero spin polarisation (N_α − N_β = 0).

Middle — the classical-RBM loop. The surviving determinants, plus the Hartree–Fock determinant as a mandatory anchor, form a persistent "memory." A truncated Hamiltonian is built in the memory subspace and diagonalized. Its eigenvector amplitudes are used to train an RBM, which then Gibbs-samples new candidate determinants. Those candidates pass a combined symmetry-and-novelty filter and are unioned with the memory to form the working space. Before diagonalization, unique α-strings and β-strings are pooled and cross-multiplied (spin-string proliferation), which brings in determinants nobody actually sampled. After diagonalization, the top-K determinants whose squared CI coefficients exceed the threshold τ_sv, and which are not already in memory, are permanently appended. The loop repeats for n iterations.

Outermost — DMET. The full system is treated at Hartree–Fock in a localized orbital basis, then split into fragments, each dressed with bath orbitals that carry its entanglement with the environment. Each fragment Hamiltonian is handed to the QSCI-RBM solver, which returns a fragment energy and electron count. A global chemical potential μ_glob is tuned by a Newton–Secant root-find until the total embedded electron count matches the true electron count within ε_con.

Baselines compared. Three solvers were run against the same DMET loop: DMET-SQD with the heuristic cap ε_spb = √|S|/2 (chosen because spin-string proliferation can inflate the subspace far beyond what was sampled, and this cap guarantees the worst-case proliferated space does not exceed |S|); DMET-SQD with ε_spb = 10⁸ (effectively unbounded, the most favourable case for SQD); and DMET-QSCI-RBM. Because the first two Newton–Secant seed points are fixed at μ_chem = 0.0 and 10⁻⁴ regardless of solver, the raw hardware bitstrings at those two points were reused across all three configurations.

Software and error mitigation. Tangelo v0.4.3 (DMET embedding and μ loop), ffsim v0.10.0 (LUCJ circuits), Qiskit v1.4.2 with Qiskit IBM Runtime v0.36.1 (execution), PySCF v2.10.0 (Hartree–Fock, CCSD, SCI), PyTorch v2.10.0 (RBM), scipy v1.15.3 (root-finding). Small-molecule validation runs used both dynamical decoupling (XpXm) and Pauli twirling — 200 randomizations at 100 shots for C₂H₄ and 700 randomizations at 64 shots for CH₅NO, active-circuit strategy — because single-fragment runs could absorb the QPU overhead. DMET runs used 100,000 shots per fragment per iteration.

Why This Matters

Impact on research. The paper reframes subspace construction, not Hamiltonian simulation, as the bottleneck of QSCI/SQD. Its central claim is that a learned generator trained directly on hardware samples can outperform both symmetry-only configuration recovery and (by design) perturbatively seeded approaches, producing much smaller diagonalization problems at equal or better accuracy. It also supplies a practical engineering bound (ε_spb = √|S|/2) for controlling spin-string proliferation — a step the authors argue can silently inflate SQD subspaces well beyond what was sampled.

Real-world applications (drawn from the paper's own framing):

• Structure-based drug design — the concrete target is the covalent inhibitor Carmofur bound to the SARS-CoV-2 main protease.
• Pharmaceutical-scale quantum chemistry of binding pockets, where full-system exact diagonalization is described as orders of magnitude beyond exascale reach.
• Strongly correlated transition-metal clusters and catalytic centres, cited as prior SQD targets that classical methods struggle with.
• Materials and open-shell / excited-state problems, which the introduction lists as extensions of the QSCI/SQD paradigm that could inherit this compact-subspace generation.

Industry relevance. Every reduction in subspace size directly reduces classical diagonalization cost, which the paper identifies as the practical limiter. The method is built entirely on production tools (Qiskit, IBM Quantum hardware, PySCF, PyTorch), making it adoptable by groups already running SQD workflows, and it targets the pharma and quantum-computing vendor segments specifically.

Future Directions

• Systematic hyperparameter optimisation. The authors state explicitly that the RBM learning rate, hidden-unit count, Gibbs chain length, and singular-value threshold were set empirically, and expect grid search or Bayesian optimisation to yield more compact subspaces, faster convergence, and potentially sub-mHa accuracy at earlier iterations.

• Establish whether the high excitation orders matter. The paper observes substantial shot density at excitation orders roughly four through six relative to Hartree–Fock but leaves open whether those configurations contribute meaningfully to the converged ground-state wavefunction.

• Deeper or adaptive fragmentation. The 11-fragment partition was placed manually along single bonds; the authors acknowledge that more refined schemes respecting full bond topology and molecular geometry exist and note that the manual partition was used only to benchmark the framework, leaving improved partitioning as an open route.

• Full self-consistency and scale-up. The DMET-QSCI-RBM results are reported at a halted μ-self-consistency point, so completing the chemical-potential loop and extending the workflow to larger or multi-fragment biological targets remain natural next steps.

Target Audience

Researchers and graduate students in quantum chemistry, quantum computing, and scientific machine learning who already understand second quantization and embedding theory — particularly groups building QSCI/SQD workflows or hybrid quantum-classical pipelines for biological and strongly correlated systems. Readers interested in generative models for configuration-space search will find the equal-weighting design choice and the spin-string proliferation mechanics directly reusable. A reader wanting only the headline claim can read the abstract: roughly 4% of the subspace at chemical accuracy versus up to 20% without reaching it for the SQD baseline.

Authors’ abstract

Sample-based Quantum Diagonalization (SQD), an extension of Quantum Selected Configuration Interaction (QSCI), has emerged as a promising hybrid quantum-classical paradigm for computing molecular ground state energies. By leveraging quantum sampling instead of variational optimization, QSCI avoids barren plateaus and enables direct reconstruction of correlated electronic wavefunctions. However, existing configuration recovery techniques primarily enforce symmetry constraints without guaranteeing optimal selection of the most physically relevant configurations, often leading to unnecessarily large subspaces and increased classical diagonalization costs. In this work, we introduce a machine-learned compact subspace generation protocol based on Restricted Boltzmann Machines (RBMs), termed QSCI-RBM, and integrate it within the Density Matrix Embedding Theory (DMET) framework. The RBM is trained on quantum-sampled configurations to learn the underlying probability distribution of dominant determinants, enabling the targeted generation of high-probability configurations. We apply this framework to the simulation of a protein-ligand complex involving the inhibitor Carmofur bound to the SARS-CoV-2 main protease ($M^{\text{pro}}$). Our results demonstrate that DMET-QSCI-RBM achieves energies within the chemical accuracy threshold by accessing only approximately 4% of the configuration subspace. In contrast, standard DMET-SQD simulations failed to reach chemical accuracy while accessing up to 20% of the subspace, even as the chemical potential itself nearly converged. These findings highlight that RBM-assisted configuration generation produces significantly more compact subspaces while preserving physical accuracy, thereby reducing classical computational overhead and enabling the scalable quantum embedding simulation of complex biological systems.