Skip to main content
Back to timeline
arXivSource publication:

SpinAware reads one mean-field solution per spin state and flags 99% of correlated orbitals, carrying transition-metal spin gaps onto IBM hardware

Synopsis

Multireference calculations on transition-metal complexes depend on choosing an active space, yet measuring which orbitals are strongly entangled requires the very correlated calculation the space is meant to make affordable. SpinAware predicts DMRG single-orbital entropies and orbital-pair mutual information from the unrestricted Hartree–Fock solution of each spin state, using 60 orbital descriptors and a transformer. Over 11 spin states it reproduces the entropy profiles with a mean absolute error of 0.027, and a candidate pool covers 99% of the DMRG-active orbitals while keeping on average 15 of 40, supporting sample-based quantum diagonalization on IBM processors up to 32 qubits.

Source-provided article image: Machine-learned active spaces enable quantum calculations of spin-state energetics of transition-metal complexes on quantum hardware
FIG. 1 · arXiv · Page 3

Interpretation

Choosing an active space no longer has to begin with a correlated calculation over a large orbital space: a machine-learning model estimates orbital entanglement directly from each spin state's mean-field solution and builds active spaces for classical and quantum multireference calculations. Entanglement-based selection previously required a preliminary DMRG calculation over a large orbital space, while projection schemes relied on a predefined list of atomic orbitals and did not distinguish spin states of the same complex. Over 11 spin states the predicted entropy profiles match DMRG with a mean absolute error of 0.027, and a candidate pool built from predicted entropies and mutual information contains 82 of the 83 DMRG-active orbitals, 99%, while holding on average only 15.2 of the 40 window orbitals per state.

The model scores each orbital in the context of all frontier orbitals, so it can separate the high- and low-spin states of one framework and identify which orbitals are correlated with each other. A network that scores each orbital independently never sees the features of its partner and cannot retain or discard two orbitals that matter only together. The input is a window of 20 occupied and 20 virtual orbitals, each described by 60 descriptors from the UHF solution, including the α–β occupation difference, the α–β energy splitting, continuous metal–ligand composition weights and exchange-coupling statistics; the model was trained on 3454 spin-pure DMRG calculations on mononuclear Cr–Ni complexes and evaluated on experimentally characterized complexes whose ligands, binding topologies, oxidation states and metals are absent from training.

The selected spaces are carried to IBM quantum processors, where sample-based quantum diagonalization solves the same active-space Hamiltonians. A quantum result can be no better than the active space fixed beforehand, and the earlier cobalt calculation fixed that space by projection onto preselected atomic orbitals. For a nickel and a cobalt complex, SqDRIFT samples reproduce the spin-state gaps of the exact CASCI calculation to within 0.10 kcal mol−1 and energies to within 0.3 kcal mol−1 for up to 32 qubits; all these spaces remain within reach of exact diagonalization, so the calculation validates the workflow rather than showing a quantum advantage.

Perspective

Results refer to the def2-SVP basis, gas-phase species at fixed geometries and elements up to Kr, and the model works in a fixed 40-orbital window around the frontier; benchmark complexes come from experimental crystal structures and were computed as the isolated metal-containing species.

Strong static correlation is compressed, as in the vanadium(III) singlets where DMRG entropies near 0.6 are predicted at about 0.15; in weakly correlated high-spin states the entropy profile is flat near any threshold, so a hard cut is ill-conditioned; singly occupied d orbitals are included by a chemical rule rather than by the model; several intermediate-spin states have no UHF reference meeting the spin-contamination criterion and fall outside the workflow; and whether the faster convergence from hardware noise persists for larger active spaces or other noise levels remains open.