Demonstration of efficient predictive surrogates for large-scale quantum processors

Researchers introduce two provably efficient classical learning models called predictive surrogates (h_cs and h_qs) that emulate the mean-value behavior of noisy quantum processors without requiring repeated quantum hardware access. Demonstrated on a 42-qubit superconducting processor, these surrogates enable pre-training of variational quantum eigensolvers (VQEs) for transverse-field Ising models using only 0.023% of the quantum measurements needed by conventional VQE, while achieving better accuracy. A second surrogate successfully identifies non-equilibrium Floquet symmetry-protected topological phase transitions classically. Both surrogates scale polynomially with qubit count and circuit depth under defined noise conditions, offering a practical pathway to broaden access to large-scale quantum processors.