IonQ and partners unveil generative AI for quantum circuits

On 16 September 2026, IonQ detailed joint research with Oak Ridge National Laboratory, Nvidia, and the University of Tennessee, Knoxville. The study shows that a trained generative transformer can directly design quantum optimization circuits, eliminating the repetitive parameter-tuning loops that traditionally make large-scale hybrid quantum computing costly to run.

In hybrid quantum optimization, problems are broken into smaller parts that require tailored circuits. In benchmark tests on a dense problem with 100 decision variables, prior methods saw circuit-finding times climb from about 34 seconds on 4 qubits to more than 11 minutes on 12 qubits. The generative model, dubbed DQAOA-GPT, kept runtimes near 28 seconds across every tested size while roughly doubling solution quality as subproblems grew.

The research was led by Oak Ridge National Laboratory and simulated on an Nvidia H200 GPU inside the Defiant2 system using the Nvidia CUDA-Q platform and cuQuantum SDK. For each subproblem, the generative model sampled ten candidate circuits, scored each simulation, and applied the best candidate to update the overall solution.

The paper won a best paper award at IEEE Quantum Week 2026 in Toronto, where IonQ had nine accepted papers. IonQ stated that the method provides a path to scale optimization workflows across its existing and future hardware generations.