MAIN RESULTS

Neural network on a quantum controller

Quantum Machines released the OPX1000 controller in 2023 and made it available to the consortium. Its FPGA can host small feedforward networks, but the recurrent networks required for non-Markovian feedback exceed the available capacity. Quantum Machines and NVIDIA therefore developed the DGX Quantum system, which couples the OPX1000 to a Grace–Hopper GPU through a deterministic low-latency interface. The system was delivered to ENS de Lyon in June 2025 as one of its early end-user sites. An adaptive Bayesian T₁ tracking experiment measured a non-wait classical overhead of about 11 µs per shot, short compared with the cavity coherence time and with the qubit reset time. The duty cycle of the closed loop is thus limited by the physics rather than by the software stack.

Neural network for state preparation

A neural network was trained in simulation to output directly the four control fields (qubit and cavity I/Q) that prepare any Schrödinger cat state |α⟩ + e^iφ|−α⟩ in a 3D microwave cavity dispersively coupled to a transmon qubit. Once trained, the network produces a new pulse sequence for a new (α, φ) in microseconds, about five orders of magnitude faster than an iterative GRAPE optimization for each new state. The experiment at ENS de Lyon reached fidelities close to those obtained by per-state GRAPE optimization (PRX Quantum 6, 010321 (2025)).

Stabilization of a quantum state by measurement-based feedback

The Erlangen team extended the Feedback-GRAPE method to long recurrent control sequences and released the open-source Python package feedback-grape, with GPU acceleration, JIT compilation and automatic differentiation based on JAX. Simulations show that a recurrent network with about 30 000 parameters can stabilize the superposition |2⟩ + |3⟩ of a cavity against decay, using only stroboscopic projective measurements of a dispersively coupled qubit. The same approach applied to the Gottesman–Kitaev–Preskill code improves the logical lifetime by about 100% relative to the standard small-BIG-small protocol (Phys. Rev. Lett. 134, 020601 (2025)). On the experimental side, ENS de Lyon demonstrated single-shot photocounting and photon-number tracking in a cavity whose decay rate is four orders of magnitude smaller than both the dispersive coupling and the qubit emission rate, thanks to a custom notch filter and a pogo-pin galvanic contact (Phys. Rev. Lett. 133, 153602 (2024)). Quantum jumps are observed as photons leave the cavity one at a time. A new reset scheme for long-lived cavities with small static cross-Kerr, the Differential-Drive Reset, was discovered and characterized in the process. The deployment of the trained networks on the actual cavity using the DGX Quantum is ongoing.

Quantum error correction on a cat code

The initial route to the four-photon dissipation needed to stabilize a four-component cat code relied on hard pumping of a nonlinearity derived from the ATS circuit. Experiments showed that at the required pump powers, parasitic effects (qubit ionization, dressed-state populations, pump-induced decoherence) become dominant before the four-photon channel does. ENS de Lyon and Alice & Bob therefore designed the autoparametric cat scheme, which reaches strong multiphoton coupling by structural design rather than by external pumping (Phys. Rev. X 14, 021019 (2024)). The experimental demonstration of four-component cat-code stabilization with this scheme continues beyond the formal end of the project.

Cloud access and technology transfer

The DGX Quantum / OPX1000 product is now commercialized by Quantum Machines. Alice & Bob opened cloud access to its cat-qubit hardware through the Felis Cloud platform, which constitutes the first cat qubit accessible on the cloud. The neural-network-driven control layer on top of it is the subject of ongoing work. The project trained one PhD student to graduation, two postdoctoral researchers and several master students, and included two European deep-tech companies whose roadmaps were shaped by its results.