Neural network designed pulse sequences for robust control of single-triplet qubits

摘要

In the near future, more and more laborious tasks will be replaced bymachines. In the context of quantum control, the question is, can machinesreplace human beings to design reliable quantum control methods? Here weinvestigate the performance of machine learning in composing composite pulsesequences which are indispensable for a universal control of singlet-tripletspin qubits. Subject to the control constraints, one can in principle constructa sequence of composite pulses to achieve an arbitrary rotation of asinglet-triplet spin qubits. In absence of noise, they are required to performarbitrary single-qubit operations due to the special control constraint of asinglet-triplet qubit; Furthermore, even in a noisy environment, it is possibleto develop sophisticated pulse sequences to dynamically compensate the errors.However, tailoring these sequences is in general a resource-consuming process,where a numerical search for the solution of certain non-linear equations isrequired. Here we demonstrate that these composite-pulse sequences can beefficiently generated by a well-trained, double-layer neural network. Forsequences designed for the noise-free case, the trained neural network iscapable of producing almost exactly the same pulses developed in theliterature. For more complicated noise-correcting sequences, the neural networkproduced pulses with a slightly different line-shape, but the robustnessagainst noises remains about the same. These results indicate that the neuralnetwork can be a judicious and powerful alternative to existing techniques, indeveloping pulse sequences for universal fault-tolerant quantum computation.