A team proposes a quantum-centric supercomputing (QCSC) framework that breaks complex molecular problems into parallel sub-problems, each requiring fewer qubits, shallower circuits and a smaller shot budget. The idea addresses one of the practical limits of current QCSC approaches, such as sample-based quantum diagonalization (SQD), which is held back in large active spaces by qubit counts, circuit depth and the cost of Hamiltonian diagonalization.

The global solution is rebuilt in two steps. First, a bit assignment strategy optimises a cost function to map local bit indices onto global ones, yielding a mapped tensor product state (MTPS). Then an inter-fragment correlation recovery module restores the cross-fragment interactions lost when the problem is split. The method uses a multi-layered generative learning architecture: it builds on the recently proposed physics-informed generative SQD (PIGen-SQD) as a distributed QCSC solver, and adds a second generative module that models configuration distributions and identifies the dominant sectors of the global Hilbert space.

For water and formamide dimers, the approach achieves up to a 95% reduction in the diagonalization dimension while improving energy accuracy by an order of magnitude over conventional SQD. Applied to the water hexamer (H₂O)₆ in its cage configuration, the technique successfully decomposes a 120-qubit problem into six 20-qubit monomeric subsystems and stitches the monomers together under a dimeric approximation, showing excellent agreement with classical reference methods.

The result points to a route for tackling larger molecular systems without waiting for quantum processors with many more qubits, exploiting instead the coordination between the quantum processor and classical supercomputers. How the framework scales to larger active spaces, and what errors the dimeric approximation introduces when stitching fragments in more extended hydrogen-bonded systems, remain open questions.