Machine learning for efficient generation of universal hybrid quantum computing resources

TL;DR

The paper demonstrates that deep reinforcement learning can optimally control a time-multiplexed, measurement-based optical circuit to generate squeezed cat states with an average success rate of $98\%$, providing a viable path to GKP precursors for fault-tolerant continuous-variable quantum computing. By modeling the circuit as a Markov decision process and using PPO, the authors extract state-generation strategies that are robust to measurement outcomes and leverage PNR detection. In addition to high-performance multi-step generation, they identify fixed-parameter single-step protocols, including a moving-target analysis that yields near-optimal fidelities and a practical GKP-breeding pathway achieving $0.99$ fidelity to standard GKP states. These results highlight a scalable, qubit-free optical route to non-Gaussian resources, with immediate implications for efficient resource generation in photonic quantum computing.

Abstract

We present numerical simulations of deep reinforcement learning on a measurement-based quantum processor--a time-multiplexed optical circuit sampled by photon-number-resolving detection--and find it generates squeezed cat states with an average success rate of 98%, far outperforming all other similar proposals.

Figures (8)

• Figure 1: Quantum optical circuit for squeezed cat state generation driven by deep reinforcement learning. The agent's input is shown in red and its parameters in green. The end result is in blue. VBS: variable beam-splitter; PNR: photon-number-resolving (measurement).
• Figure 2: Wigner functions of the 4 target squeezed cat states ($r$=1.38, $\alpha$=3) used in the reward function calculation of Eq. (\ref{['eq:f']}).
• Figure 3: Histograms for 1250 generation episodes ($m=50$).(a), output state fidelity; (b), detect photon number per episode; (c), steps per episode, with resets; (d), steps per episode, between resets.
• Figure 4: Fidelities of the output state of the first step with a target squeezed cat state with $(\alpha,r)=(3,1.38)$. The vertical dot-dashed line marks the initial transmittivity $\tau_1^2$=0.12 ($\tau_1$=0.34).
• Figure 5: Average fidelity versus transmissivity. The dot-dashed line marks the initial transmissivity ($\tau=0.34$) chosen by the trained agent.