Neural-network-based design and implementation of fast and robust quantum gates
Marko Kuzmanović, Ilya N. Moskalenko, Yu‐Han Chang, Ognjen Stanisavljević, Christopher Warren, Emil Hogedal, Anuj Aggarwal, Irshad Ahmad, Janka Biznárová, Mamta Dahiya, Marcus Rommel, Andreas Nylander, Giovanna Tancredi, Gheorghe Sorin Paraoanu
We present a continuous-time, neural-network-based approach to optimal control in quantum systems, with a focus on pulse engineering for quantum gates. Leveraging the framework of neural ordinary differential equations (neural ODEs), we construct control fields as outputs of trainable neural networks, thereby eliminating the need for discrete parametrization or predefined bases. This allows for the generation of smooth, control-hardware-agnostic pulses that can be optimized directly using differentiable integrators. As a case study, we design and experimentally implement a short and detuning-robust π / 2 pulse for photon number parity measurements in superconducting transmon circuits. This is achieved through simultaneous optimization for robustness and suppressing the leakage outside of the computational basis. These pulses have very high fidelity ( ≈ 99.95 % ) and realize the parity measurement with an efficiency greater than 99.9% over a detuning range of ≈ ± 20 MHz , thereby outperforming traditional techniques while retaining comparable gate durations. This showcases the potential of neural ODEs for high-performance quantum control in experimentally relevant settings.
We present a continuous-time, neural-network-based approach to optimal control in quantum systems, with a focus on pulse engineering for quantum gates. Leveraging the framework of neural ordinary differential equations (neural ODEs), we construct control fields as outputs of trainable neural networks, thereby eliminating the need for discrete parametrization or predefined bases. This allows for the generation of smooth, control-hardware-agnostic pulses that can be optimized directly using differentiable integrators. As a case study, we design and experimentally implement a short and detuning-robust π / 2 pulse for photon number parity measurements in superconducting transmon circuits. This is achieved through simultaneous optimization for robustness and suppressing the leakage outside of the computational basis. These pulses have very high fidelity ( ≈ 99.95 % ) and realize the parity measurement with an efficiency greater than 99.9% over a detuning range of ≈ ± 20 MHz , thereby outperforming traditional techniques while retaining comparable gate durations. This showcases the potential of neural ODEs for high-performance quantum control in experimentally relevant settings.
The presence of decoherence in quantum computers necessitates the suppression of noise. Dynamically corrected gates via specially designed control pulses offer a path forward, but hardware-specific experimental constraints can cause complications. Existing methods to obtain smooth pulses are either restricted to two-level systems, require an optimization over noise realizations, or limited to piecewise-continuous pulse sequences. In this paper, we present the first general method for obtaining truly smooth pulses that minimizes sensitivity to noise, eliminating the need for sampling over noise realizations and making assumptions regarding the underlying statistics of the experimental noise. We parametrize the Hamiltonian using a neural network, which allows the use of a large number of optimization parameters to adequately explore the functional control space. We demonstrate the capability of our approach by finding smooth shapes, which suppress the effects of noise within the logical subspace as well as leakage out of that subspace.
Real-time adaptive control of quantum systems requires rapid generation of robust, high-fidelity pulses across a continuous range of operating conditions. Standard optimization algorithms such as gradient-ascent pulse engineering (GRAPE) solve each instance independently, discarding information between runs and requiring costly reinitialization when parameters change. We present an approach to robust optimal quantum control based on model-based reinforcement learning, in which a single neural network -- embedding the Hamiltonian directly into the training pipeline -- generates robust gates across an entire family of gate configurations, without pre-computed training data. Demonstrated on a single-spin (two-level) system, the trained networks produce pulses for arbitrary rotation angles over a range of pulse durations, detunings, and field inhomogeneities in milliseconds, at fidelities comparable to multi-seed GRAPE. The framework is inherently adaptable: any parameter entering the Hamiltonian can serve as a network input, extending the approach to different systems and control settings. Beyond speed, the network reveals structure in the control landscape: it discovers the same structured phase profiles that appear in GRAPE solutions -- made identifiable through fidelity-invariant symmetry transformations -- but more consistently than independent optimization. This consistency enables smooth interpolation across the entire trained parameter space.
The potential of quantum computing is fundamentally constrained by the inherent susceptibility of qubits to noise and crosstalk, particularly during multi-qubit gate operations.
Existing strategies, such as hardware isolation and dynamical decoupling, face limitations in scalability, experimental feasibility, and robustness against complex noise sources.
In this manuscript, we propose a physics-guided neural control (PGNC) framework to generate robust control pulses for superconducting transmon qubit systems, specifically targeting crosstalk mitigation.
By combining a hardware aware parameterization with a Hamiltonian-informed objective that accounts for condition-dependent crosstalk distortions, PGNC steers the search toward smooth and physically realizable pulses while efficiently exploring high dimensional control landscapes.
Numerical simulations for the CZ gate demonstrate superior fidelity and pulse smoothness compared to a Krotov baseline under matched constraints.
Taken together, the results show consistent and practically meaningful improvements in both nominal and perturbed conditions, with pronounced gains in worst-case fidelity, supporting PGNC as a viable route to robust control on near-term transmon devices.
Fault-tolerant quantum computing requires high-fidelity quantum gates that are resilient to control errors and device imperfections. Here, we develop energy-optimal control pulses for single-qubit gates that are simultaneously robust against frequency detuning and waveform inhomogeneities while also suppressing leakage outside the computational subspace. Combining the optimized pulse envelopes with an adaptive control scheme, we experimentally demonstrate optimal robust gates on a superconducting quantum processor. For a gate duration of 90ns, we achieve fidelities above 99.9% for the X gate and 99.6% for Hadamard-type gates. For the X gate, we maintain fidelities above 99.5% for simultaneous drive-amplitude variations up to 20% and qubit detuning up to 4 MHz. This robustness outperforms that of conventional Gaussian pulses by an order of magnitude, without much difference in peak gate fidelity. These results establish a route towards robust single-qubit control in large-scale quantum processors, advancing the development of fault-tolerant quantum computing.
In this work, we first consider standard quantum control problems for superconducting transmon qubits. Such quantum systems have been studied extensively both from an analytical and a numerical perspective [MSG&al11; TMW16]. Analytical pulses usually show dependencies from system parameters that are not easily found in numerical solutions. Therefore, our approach is then to employ neural networks to learn the functional dependence of optimal quantum control solutions from system parameters. We show that we can optimize such solutions analytically for single and two-qubit gates using either very few parameter samples or large numbers of pulse frequencies for large parameter ranges. Afterward, we move to a higher level of optimization. We consider variational gates in trapped-ion quantum computers [MMN&al16]. In this setting, we employ a hybrid scheme that uses both reinforcement learning and continuous optimization to optimize both circuit structure and variational angles concurrently [BDS&al18; SEL&al22]. We show that our reinforcement learning algorithm assisted by a continuous optimizer can construct effective solutions to the gate synthesis problem that matches and surpasses standard circuit compilers. Next, we consider variational optimization of non-linear maps [HCS&al23] acting on quantum states such as entanglement purification protocols. An entanglement purification protocol acts as a highly non-linear map, but it still outputs a faithful representation of a quantum system. Optimizing it for a specific family of states requires reducing the number of operations needed to retrieve the original state. This optimization has the potential to make purification protocols easier to implement on a quantum device. In fact, the exponential scaling of the purification protocol with the number of states [DB07] implies that such protocols may not be applied directly without careful engineering. We also show that the performance of standard purification protocols for arbitrary two-qubit input states leads to poor output values of the concurrence. Our optimized protocols prove instead able to increase the value of the concurrence above the maximum limit of traditional analytical protocols. We also show how the twirling operation becomes an obstacle to the performance of the protocol itself when considering random two-qubit states, although it is a useful tool in the design of entanglement purification protocols. Finally, we study parameter sampling in quantum circuits, focusing in particular on the LCU methods. Such problems are particularly interesting for meta-variational settings [CKA21] where we compute the average over observables evaluated at different points in the parameter space. They are also relevant for optimal quantum control algorithms because the computation of the fidelity with respect to a target operation is the basis of most optimal quantum control routines. In conclusion, we analyze several optimization problems that are relevant for quantum science and technology. We show that machine learning-assisted solutions can be applied successfully to engineer optimal quantum control pulses and compilation strategies based on variational angles. Also in the context of entanglement purification, we show how our optimized protocols can surpass current strategies for multiparametric families of states.
Quantum gate design is often represented as pulse optimization, although the physical object that implements a gate is the full controlled evolution generated by the pulse. Here we use physics-informed neural networks to represent single-qubit gate design at this evolution level: the control fields, the Bloch-state trajectories, and the total duration are learned together under the Bloch equation. This changes the optimized object from pulse amplitudes to a differentiable physical process whose structure can be inspected and refined. For rotation gates, the optimized evolutions recover the physical organization expected for bounded single-qubit control, with no prescribed pulse ansatz or duration scan. For a geometric gate, the representation identifies localized bottlenecks in maintaining the geometric condition and turns this diagnosis into feedback, reducing the residual path error while preserving high fidelity. Thus physics-informed learning is used not only to synthesize gates, but also to make optimized quantum controls physically readable, diagnosable, and locally refinable. This process-level view may be especially useful for adapting gates to hardware-specific, task-specific, and locally varying experimental constraints.