QBM: A Quantum Bayesian Machine for Relaxed Energy Prediction on Metal-Oxide Catalysts

Whether quantum circuits add useful search or representation to catalyst-energy surrogates remains unclear under matched classical comparisons. We evaluate a Quantum Bayesian Machine (QBM) on the OC22 relaxed-total-energy task, holding structural descriptors and data partitions fixed while varying subset proposal, fidelity-kernel regression and variational residual learning. In the recorded CPU-simulated workflow, the circuit-assisted selector fitted 333 candidates versus 65,535 under enumeration, with mixed validation-error changes; original candidate lists and timing logs are unavailable. A selected fidelity-kernel setting lowered in-distribution mean absolute error by a median 0.00433 eV per atom against paired linear ridge across six fits, but showed no median out-of-distribution gain. Re-uploading had a median error only 0.00128 eV per atom below ridge across five seeds with overlapping ranges, and its residual-target construction is incompletely archived. These results delimit component behaviour under ideal simulation and identify the controls and provenance needed to test whether circuit-specific benefits persist.