Quantum computing requires the optimization of control pulses to achieve high-fidelity quantum gates. We propose a machine learning-based protocol to address the challenges of evaluating gradients and modeling complex system dynamics. By training a recurrent neural network (RNN) to predict qubit behavior, our approach enables efficient gradient-based pulse optimization without the need for a detailed system model. First, we sample qubit dynamics using random control pulses with weak prior assumptions. We then train the RNN on the system's observed responses, and use the trained model to optimize high-fidelity control pulses. We demonstrate the effectiveness of this approach through simulations on a single
ST0 qubit.
How to implement multi-qubit gates efficiently with high precision is essential for realizing universal fault-tolerant computing. For a physical system with some external controllable parameters, it is a great challenge to control the time dependence of these parameters to achieve a target multi-qubit gate efficiently and precisely. Here we construct a dueling double deep Q-learning neural network (DDDQN) to find out the optimized time dependence of controllable parameters to implement two typical quantum gates: a single-qubit Hadamard gate and a two-qubit CNOT gate. Compared with traditional optimal control methods, this deep reinforcement learning method can realize efficient and precise gate control without requiring any gradient information during the learning process. This work attempts to pave the way to investigate more quantum control problems with deep reinforcement learning techniques.
In this paper, we present a machine learning framework to design high-fidelity multi-qubit gates for quantum processors based on quantum dots in silicon, with qubits encoded in the spin of single electrons. In this hardware architecture, the control landscape is vast and complex, so we use the deep reinforcement learning method to design optimal control pulses to achieve high fidelity multi-qubit gates. In our learning model, a simulator models the physical system of quantum dots and performs the time evolution of the system, and a deep neural network serves as the function approximator to learn the control policy. We evolve the Hamiltonian in the full state-space of the system, and enforce realistic constraints to ensure experimental feasibility.
We propose a physics-informed neural network (PINN) approach for high-fidelity quantum gate learning in both one- and two-qubit systems. We train a PINN using the time-dependent Schrödinger equation, where smooth control fields with fixed boundary conditions are obtained by minimizing a loss function that considers model dynamics, gate error, physical constraints, regularization terms, and multiple initial conditions. Training considers pairs of mutually unbiased bases, while independent validation combines Haar-random input states with the average gate fidelity. Unitary simulations yield an average gate infidelity of
8.49×10−8 for Hadamard and values between
3.39×10−7 and
4.96×10−6 for controlled-NOT under
XX,
YY, and
ZZ interactions. We examine how the learned control fields depend on the coupling, training configuration, and available local fields, including five exchange-type dipolar scenarios. We test the learned control fields under the Lindblad master equation with dephasing and amplitude damping, theoretically reproducing the expected first-order fidelity loss in the weak-dissipation regime. Finally, boundary-adapted Chebyshev expansions represent 139 learned controls with a minimum common count of
K∗=24 coefficients per control for all eleven family-averaged mean squared errors to fall below
10−6. Separately, square-pulse and Strang propagation at
K=30 preserve all 31 continuous polynomial references within infidelity
10−5 at
N=128 and all larger tested resolutions, under both sampling rules. These results connect multi-state gate synthesis, control-resource characterization, and compact square-pulse playback.
Quantum optimal control theory (QOCT) typically addresses the control of physical qubits. Finding explicit pulse control sequences in such a framework is challenging, especially when an underlying physical model is unknown. We propose a deep reinforcement learning (DRL) method, which doesn't require any underlying gate model or qubit pre-calibration, capable of controlling a superconductive qubit via analog pulses acting in the IBM Qiskit Pulse environment. We applied the method to build a single-qubit gate with high fidelity and short duration at pulse level. In particular, the DRL agent approximated the X90 gate at the physical layer on the IBM Armonk transmon superconductive qubit simulated by the Qiskit Pulse simulator. The learned sequence has an average gate fidelity greater than 0.978 and a duration of 58 ns only, faster than the default X90 pulse IBM implementation, which has a runtime of 140 ns. Without prior knowledge and gate model knowledge, the agent learned a non-traditional shaped microwave pulse, providing an alternative strategy for controlling noisy quantum states.
We present a general method to quickly generate high-fidelity control pulses for any continuously-parameterized set of quantum gates after calibrating a small number of reference pulses. We find that interpolating between optimized control pulses for different quantum operations does not immediately yield a high-fidelity intermediate operation. To solve this problem, we propose a method to optimize control pulses specifically to provide good interpolations. We pick several reference operations in the gate family of interest and optimize pulses that implement these operations, then iteratively re-optimize the pulses to guide their shapes to be similar for operations that are closely related. Once this set of reference pulses is calibrated, we can use a straightforward linear interpolation method to instantly obtain high-fidelity pulses for arbitrary gates in the continuous operation space. We demonstrate this procedure on the three-parameter Cartan decomposition of two-qubit gates to obtain control pulses for any arbitrary two-qubit gate (up to single-qubit operations) with consistently high fidelity. Compared to previous neural network approaches, the method is 7.7x more computationally efficient to calibrate the pulse space for the set of all single-qubit gates. Our technique generalizes to any number of gate parameters and could easily be used with advanced pulse optimization algorithms to allow for better translation from simulation to experiment.