IBM, Mitsubishi Chemical, and Keio Train Generative Quantum Eigensolvers on 70-Qubit Hardware

A research consortium comprising IBM Quantum, Mitsubishi Chemical Corporation, Keio University, and NVIDIA has demonstrated an in-the-loop quantum hardware training framework for autoregressive circuit generators up to 70 qubits. Published on arXiv, the study used the 156-qubit IBM Heron processor ibm_kobe to train transformer-based Generative Quantum Eigensolver (GQE) policies, which were subsequently applied zero-shot to simulate complex organotin photoresist molecules for extreme ultraviolet (EUV) lithography.

Simulating strongly correlated electrons in organometallic clusters presents an exponential barrier for classical configuration interaction methods, while deep parameterized circuits quickly decohere on physical quantum processors. Under Sample-based Quantum Diagonalization (SQD), the quantum processor only samples salient bitstring configurations, after which a classical processor reconstructs the active subspace and diagonalizes the molecular Hamiltonian. State preparation quality dictates recovery accuracy, but conventional ansätze often generate prohibitive gate depths once mapped to physical hardware topologies.

To eliminate manual circuit synthesis, the team implemented a two-stage reinforcement learning curriculum. After initial pre-training on a noiseless simulator across transition-metal complexes, training continued directly on the ibm_kobe processor across 46- and 70-qubit active spaces. The Stage-2 training loop incorporated a specialized reward penalty to prevent policy collapse onto the classical Hartree–Fock state, consuming 120 million physical hardware shots across 600 calibration jobs.

The resulting policies were deployed without retraining against 11 active-space tiers of methyltin trihydroxide, Sn(OH)₃-Me, and its mono- and di-hydrates. In operational inference, the policy proposes ten candidate circuits that are executed in a single quantum batch job. A classical ranking pass evaluates configuration counts to select the optimal circuit, after which subspace recovery and Hamiltonian diagonalization proceed entirely on local GPUs without requiring additional quantum execution time.

The research was supported in part by Japan’s Cross-ministerial Strategic Innovation Promotion Program (SIP) and the RIKEN JHPC-quantum initiative. Full technical details are available in the preprint on arXiv here, open-source code on GitHub here, and experimental datasets on Zenodo here.

November 11, 2026