Neural Networks Accelerate Design of Superconducting Quantum Systems

A new application of deep neural networks accelerates the design of superconducting radio-frequency cavities and transmon qubits for bosonic quantum computation. Joseph Yaker of the Superconducting and Quantum Materials System Centre (SQMS), and colleagues, in collaboration with the University of Cambridge, Illinois Institute of Technology, and Northwestern University, tackle the challenge of inverse design, determining device geometries to achieve specific electromagnetic and coupling targets. Traditionally, this becomes computationally expensive as systems scale, demanding significant resources and time for even modest design explorations. Their two deep neural network approaches rapidly map desired device behaviour to candidate designs, achieving accuracy within approximately 5% for cavity observables and 2% for transmon qubit parameters including coupling rate, frequency, and anharmonicity, as verified through re-simulation. This fast alternative to iterative simulation studies represents a key step towards scalable design of complex quantum systems.

Deep learning enables two percent accuracy in transmon qubit design prediction

Transmon qubit parameters are now predicted with approximately 2% accuracy, a substantial improvement over previous methods reliant on computationally expensive iterative simulations. Conventional methods, typically involving finite element analysis and optimisation algorithms, struggled to achieve comparable accuracy within reasonable timescales, often requiring weeks or months of computation for a single design iteration. Deep neural networks directly map desired device behaviour to candidate designs, bypassing the need for lengthy trial-and-error processes, which is particularly important for scaling up quantum systems where the number of design variables increases exponentially. The inherent complexity arises from the need to simultaneously optimise multiple parameters, including qubit frequency, anharmonicity (which defines the nonlinearity of the qubit and enables quantum gate operations), and coupling strength to the resonant cavity mode.

This level of precision, confirmed by re-simulation, unlocks the rapid design of complex quantum circuits previously limited by the prohibitive cost of exploring vast design spaces. Superconducting radio-frequency cavities and transmons, essential components for bosonic quantum information processing, can now be created with unprecedented efficiency. Bosonic quantum computation leverages the quantum properties of harmonic oscillators, and superconducting circuits provide a promising platform for realising these oscillators with high coherence and controllability. Utilising deep neural networks trained on data generated from detailed simulations in COMSOL, a finite element analysis software package, superconducting radio-frequency cavities can be designed with electromagnetic properties tailored to approximately 5% accuracy. This includes parameters such as the resonant frequency, quality factor (Q-factor, representing energy loss), and mode profile, all crucial for efficient qubit coupling and information storage.

Expanding on this capability, the team predicted qubit-cavity coupling strength, qubit frequency, and anharmonicity to within 2%, a figure confirmed through re-simulation. The networks employ a layered approach with linear layers and ReLU activation functions, processing architectures with between 3 and 128 parameters. The choice of ReLU (Rectified Linear Unit) activation functions introduces non-linearity, enabling the network to learn complex relationships between input design parameters and output qubit characteristics. The range of 3 to 128 parameters within the network architecture allows for a trade-off between model complexity and computational efficiency. While these results represent a major leap in design efficiency, the current focus remains on isolated components, not yet accounting for the complex interactions and fabrication tolerances inherent in large-scale quantum systems. Future work will need to address the challenges of modelling cross-talk between qubits, the impact of material imperfections, and the limitations imposed by fabrication processes.

Accelerating quantum computer design through deep learning and reduced reliance on performance

Precise design is required to optimise the electromagnetic properties of superconducting cavities and transmon qubits, vital components in building more powerful bosonic quantum computers. The challenge of ‘inverse design’, determining the physical shape needed to achieve specific performance targets, was tackled, a process traditionally hampered by extensive computational demands. The computational burden stems from the need to solve Maxwell’s equations for complex three-dimensional geometries, requiring significant computational resources and time. Current deep learning approaches rely on re-simulation to validate performance, raising questions about how accurately these digitally-created designs will translate into real-world devices subject to manufacturing imperfections and material variations. These imperfections can introduce deviations from the simulated performance, necessitating robust design strategies that account for these uncertainties.

This represents a significant step forward in the development of complex quantum systems, as the team in Santa Barbara have demonstrated a faster route to designing these building blocks of more powerful quantum computers. The advancement shifts the focus from simulating designs to directly creating them, potentially accelerating innovation in bosonic quantum information processing. The ability to rapidly iterate on designs allows researchers to explore a wider range of possibilities and optimise performance for specific applications. Deep learning networks offer a new approach, circumventing the limitations of traditional, computationally intensive methods and enabling the rapid generation of candidate designs tailored to specific electromagnetic and coupling requirements. The potential impact extends beyond fundamental research, offering a pathway towards the development of commercially viable quantum technologies. As a result, future work can explore increasingly complex quantum architectures and fabrication techniques, building upon this foundation of rapid, intelligent design and paving the way for scalable quantum systems. This includes investigating novel qubit designs, exploring different cavity geometries, and developing automated fabrication processes to realise these designs with high precision and yield.

The researchers successfully developed deep learning networks that rapidly generate designs for superconducting radio-frequency cavities and transmon qubits. This is important because designing these components traditionally requires extensive and time-consuming computer simulations. The networks propose cavity geometries and qubit designs that meet specified electromagnetic and coupling targets to within approximately 5% and 2% respectively, as confirmed by re-simulation. The authors state that future work will focus on exploring more complex quantum architectures and fabrication techniques based on this intelligent design approach.