The Signal
IBM, RIKEN and NVIDIA report 79-fold gain in iron-sulfur simulation
A preprint reports a sample-based quantum diagonalization calculation of an iron-sulfur cluster within 4.3 millihartree of a DMRG result. IBM says that is 79 times closer than the team’s earlier closed-loop calculation on the same cluster.

A quantum-classical calculation of an iron-sulfur cluster came within 4.3 millihartree of a DMRG reference; IBM says the result was 79 times closer than the team’s earlier closed-loop calculation on the same cluster. The result appears in an arXiv preprint by researchers affiliated with IBM, RIKEN and NVIDIA, and was highlighted in an IBM post for SC26.
A different starting point
The researchers used sample-based quantum diagonalization, or SQD, to estimate the ground-state energy of an iron-sulfur cubane. The method combines samples generated by quantum circuits with classical processing: a configuration-recovery step builds a determinant subspace, which is then diagonalized on classical computing resources.
The team modified the workflow’s reference preparation to use a broken-symmetry unrestricted Hartree-Fock starting point. That approach sets up opposite spin orientations on iron centers, aiming to represent the cluster’s magnetic interactions in the initial approximation. The paper says this change lowered the mean-field reference energy by about 534 millihartree compared with a restricted Hartree-Fock starting point.
The target model contained 54 electrons in 36 orbitals, corresponding to 72 data qubits and 88 total qubits. After repeated configuration recovery using GPU nodes on RIKEN’s ROQUO system, the deepest directly calculated SQD energy was -327.235076 hartree. That was 4.3 millihartree above the finite-D=4000 DMRG value of -327.2393952 hartree.
What the comparison shows
The preprint reports that an earlier SQD calculation on the same cluster, using a restricted Hartree-Fock reference, produced energies from -326.718 to -326.723 hartree. The new result is 512 to 517 millihartree lower in energy. IBM’s separate 79-fold comparison describes how much closer the result came to the DMRG value than the team’s earlier closed-loop SQD result.
The comparison is not a controlled test of the starting reference alone. The preprint notes that the calculations differ in subspace dimension, recovery depth, sampled pool and hardware.
Calculated result and caveat
The researchers also extrapolated their results across subspace sizes. That fit gave an estimated energy of -327.244468 hartree, with an uncertainty of 1.05 millihartree; the authors say it is statistically consistent with their extrapolated DMRG estimate. This is an extrapolated estimate, distinct from the direct SQD result that remained 4.3 millihartree above the finite-D reference.
The broken-symmetry reference is not a spin eigenstate, a limitation the authors discuss in the preprint. Their reported target is the ground-state energy; they note that calculations of spin-state gaps or spectra require a spin-adapted treatment. The work also reports executions of the workflow on Fugaku, Miyabi-G and ROQUO.
Sources
Companies in this story
Related stories
- SQD Recovers FeO Model Ground States from IBM Quantum Samples
- Researchers train chemistry circuit generator on 70-qubit hardware
- IQM says Halocene upgrade path will exceed 1,000 qubits
- IBM VQE Tests Find Error Bars Missed Energy Residuals by Up to 32×
- Qunova says HI-VQE reaches chemical accuracy in an 80-qubit calculation