Optimizing cross resonance gates using recurrent neural networks

Quantum control is a powerful technique for implementing quantum circuits and designing high-fidelity quantum gates. Besides analytical and gradient-based control protocols, recently machine learning has evolved as a tool to generate control pulses for implementing high fidelity quantum gates. In this work, we demonstrate a new deep learning model based on encoder-decoder architecture using Long Short Term Memory (LSTM) units combined with convolution layers to implement high fidelity fast quantum gates. The model architectures are trained to generate optimized pulse sequences for single and two-qubit gates with infidelities in the order of 10-5 and 10-4 respectively. Furthermore, we incorporate real-world hardware limitations by incorporating pulse constraints like amplitude and bandwidth limitations. We apply these techniques to coupled transmon qubits and study the optimal sequences with attention to leakage out of computational subspace and the effect of decoherence.