AI agent QORA designs its own quantum control pulses, nearly doubling coherence time over the best human sequence in a diamond spin-1 ensemble
Synopsis
The multi-agent system QORA closes a loop between theoretical reasoning, automated calibration and measurement to search for dynamical decoupling sequences on a dense ensemble of spin-1 nitrogen-vacancy centers in diamond at 1.2 ppm. Its AID-ROD-1 sequence nearly doubles the measured spin coherence time relative to the state-of-the-art human-designed DROID-C3PO, and a second tested sequence, AID-ROD-2, reaches a comparable coherence time. Human researchers then decomposed the sequences and attributed the gain to 18-pulse blocks that suppress linear and quadratic on-site disorder and common-mode drive-amplitude errors, while arranging them over a longer cycle further reduces dipolar interactions.
Interpretation
Quantum control pulse sequences can now be discovered in a closed loop by AI agents on a real experimental apparatus, not only constructed by hand from symmetry rules. Earlier robust sequence design relied on average Hamiltonian theory and symmetry-based construction, such as DROID-C3PO, which uses a hierarchical construction to suppress disorder and control errors during finite pulses, but these rules do not uniquely fix pulse ordering and timing. On a bulk diamond NV ensemble at 1.2 ppm density, AID-ROD-1 nearly doubles the measured coherence time of DROID-C3PO, and the separately tested AID-ROD-2 reaches a comparable coherence time.
The system links sequence design to automated experiments in a feedback loop, so designs are tested directly by instrument measurements. QORA is powered by GPT-6-Astra and Sorol and connects a theoretical agent, Theo, with an experimental agent, Dex, while human scientists set goals and review experimental reports. Theo coordinates five specialized sub-agents that propose, review, simulate, classify and assemble sequences, and Dex converts selected sequences into control waveforms for an arbitrary waveform generator while monitoring 12 benchmark spin states to detect and compensate experimental drift.
The physical advantage comes from hierarchical error suppression rather than improvement in a single channel. Each 18-pulse block consists of three identical copies of a six-pulse unit, and a complete cycle contains seven such blocks, giving 126 pulses in total. Human analysis indicates that the 18-pulse blocks jointly suppress linear and quadratic on-site disorder and common-mode drive-amplitude errors, and combining the seven blocks over a longer cycle further reduces residual dipolar interactions.
Perspective
The demonstrated setting is a spin-1 NV ensemble in bulk diamond at 1.2 ppm density, where the sequences keep their advantage over DROID-C3PO as pulse spacing varies and as Rabi-frequency error is varied over a tested range. The intended readers are researchers in quantum control and AI for science.
The physical explanation comes from human researchers' subsequence analysis, in which the contribution of the quadratic disorder channel is supported jointly by spin-1 simulations and experiments in the effective spin-1/2 subspace. Extension to higher-dimensional systems beyond spin-1 remains a direction stated in the outlook.
