Title: Unlocking Autonomous Quantum Error Correction Through Deep Reinforcement Learning

Abstract:

While fault-tolerant quantum computing relies heavily on quantum error correction, conventional approaches that depend on active measurements are prone to introducing further errors. To address this, Autonomous Quantum Error Correction (AQEC) employs engineered dissipation and driving forces within bosonic systems. However, the stringent Knill-Laflamme conditions make the identification of practical encoding schemes a significant hurdle. This study leverages deep reinforcement learning, enhanced by curriculum learning, to uncover bosonic codes capable of withstanding both single-photon and double-photon losses within an approximate AQEC framework.

To expedite the reinforcement learning training process, we derive an analytical solution for the master equation under specific approximation conditions. The learning agent initially conducts a rapid exploration within a limited evolutionary timeframe to locate an encoded subspace that exceeds the breakeven point. Subsequently, it strategically refines its policy to maintain this performance advantage over longer durations. Our analysis reveals that the agent, trained in two phases, identifies the optimal codewords—specifically the Fock states $\ket{4}$ and $\ket{7}$—when accounting for both single- and double-photon loss mechanisms.

The discovered code demonstrates superior performance, surpassing the breakeven threshold across extended evolution times and achieving state-of-the-art results. Furthermore, we evaluate the code’s resilience against phase damping and amplitude damping noises. These findings underscore the promise of curriculum learning-integrated deep reinforcement learning in identifying optimal quantum error-correcting codes, particularly for early-stage fault-tolerant quantum systems.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC