Deep reinforcement learning for quantum Hamiltonian engineering

Engineering desired Hamiltonians in quantum many-body systems is essential for applications such as quantum simulation, computation and sensing. Conventional Hamiltonian engineering sequences are designed using human intuition based on perturbation theory, which may not be optimal and is unable to accommodate complex experimental imperfections. Here we search for Hamiltonian engineering sequences using deep reinforcement learning (DRL) and experimentally demonstrate that they outperform celebrated sequences in a solid-state nuclear magnetic resonance quantum simulator. We aim at decoupling strongly-interacting spin-1/2 systems and consider different experimental imperfections. The robustness of the DRL-found sequences is verified both in simulations and experiments. More interestingly, many of the DRL sequences exhibit a common pattern unknown before. By restricting the searching space to such patterns, we ultimately design sequences that are robust against dominant imperfections in our experiments. We not only demonstrate a general method for Hamiltonian engineering, but also highlight the importance of both black-box artificial intelligence and understanding of physical system in order to realize experimentally viable applications.