🔬 applied physics

A General Framework for Gradient-Based Optimization of Superconducting Quantum Circuits using Qubit Discovery as a Case Study

This paper presents a comprehensive, gradient-based optimization framework integrated with the SQcircuit software that automates the design of superconducting quantum circuits by efficiently computing eigensystem gradients, successfully demonstrating its capability to discover novel qubit architectures with superior performance metrics compared to existing designs.

Original authors: Taha Rajabzadeh, Alex Boulton-McKeehan, Sam Bonkowsky, David I. Schuster, Amir H. Safavi-Naeini

Published 2026-10-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Taha Rajabzadeh, Alex Boulton-McKeehan, Sam Bonkowsky, David I. Schuster, Amir H. Safavi-Naeini

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Building a quantum computer is like trying to tune a radio to a single, perfect station while standing in a storm of static. The devices that hold the information in these machines, called qubits, are incredibly fragile. They are built from tiny loops of superconducting wire that carry electricity without resistance, but they are so sensitive that even a whisper of heat or a stray magnetic field can scramble their data. For decades, scientists have been manually designing these circuits, tweaking the size of capacitors and the strength of magnetic junctions by hand, hoping to find a combination that keeps the qubit stable long enough to do useful work. It is a slow, trial-and-error process, much like trying to find the perfect recipe by tasting every possible variation of an ingredient. The goal is always the same: to create a qubit that can perform millions of calculations before it loses its information, all while resisting the noise of the laboratory environment.

A team of researchers at Stanford University has now developed a new way to speed up this search. Instead of guessing and checking, they built a digital framework that uses a powerful mathematical tool to automatically design better superconducting circuits. They treated the design process like a problem that could be solved by a computer learning from its own mistakes. By connecting their circuit simulation software to a system capable of calculating how tiny changes in a circuit's shape affect its performance, they created a feedback loop. The computer would propose a design, calculate how well it would work, and then immediately adjust the design to improve it, repeating this cycle thousands of times until it found a configuration that was far superior to anything humans had previously built.

The researchers applied this method to the specific task of discovering new types of qubits. They set the computer loose on a vast landscape of possible circuit shapes, ranging from simple loops to more complex arrangements with multiple junctions and capacitors. The computer was given a clear set of goals: maximize the number of operations the qubit could perform before failing, while ensuring it remained insensitive to fluctuations in magnetic fields and electric charges. It was also instructed to stay within the physical limits of what can actually be manufactured in a lab, such as the smallest and largest sizes for the components. The system explored millions of possibilities, discarding designs that were too sensitive to noise or too slow, and refining those that showed promise.

The results of this automated search were striking. The computer discovered several new circuit designs that outperformed the best qubits currently in use in terms of their theoretical potential. One of the most successful designs, which the researchers call a "JL" circuit, showed a potential to perform over a million operations, a significant jump from the hundreds of thousands achieved by existing heavy fluxonium qubits. More importantly, these new designs were much more robust against magnetic noise, a common enemy that causes qubits to lose their data. The computer found that by carefully balancing the size of the inductors and the strength of the junctions, it could create a circuit that was naturally resistant to the environmental disturbances that usually plague these devices. However, the researchers noted that the comparison of these optimized qubits to existing ones was intended solely to underscore the inherent potential of each circuit topology, rather than to claim a direct replacement for all current hardware.

Not every design the computer found was a winner, and the process revealed some interesting truths about the nature of these circuits. In some cases, the computer proposed complex structures with many parts, only to realize during the optimization that the best performance came from effectively removing some of those parts. For instance, one design that looked like a complicated hybrid of two different qubit types turned out to work best when one part of the circuit was made so large that it dominated the behavior, effectively simplifying the whole system back to a known, reliable form. This suggests that while complexity can offer new possibilities, the most effective solutions often rely on a clean, focused architecture. The researchers also found that some of the most promising new designs had properties that were not just slightly better, but an order of magnitude superior to existing technology, particularly in how long they could hold onto their quantum state, though they noted that some optimized designs ultimately reduce to or are nearly equivalent to previously discovered qubits with fewer elements.

The framework the team built is not limited to just finding better qubits. Because the system can calculate how any change in a circuit affects its performance, it can be used to solve other design problems in quantum hardware. It could help engineers create better connections between qubits, design more efficient sensors, or build circuits that are less prone to manufacturing errors. The key innovation is that the system does not need to be taught the rules of physics for every new problem; it simply needs a definition of what "good" looks like, and it will use the laws of physics to find the path to that goal. This approach shifts the burden of design from human intuition to computational power, allowing researchers to explore a much wider range of possibilities than was ever possible before.

The study confirms that the path to better quantum computers may not lie in inventing entirely new materials or exotic physical phenomena, but in optimizing the circuits we already know how to build. By letting a computer systematically explore the space of all possible designs, the researchers have uncovered configurations that human engineers might have overlooked. These new designs are not just theoretical curiosities; they are specific arrangements of capacitors, inductors, and junctions that can be built in a laboratory. The researchers have made their software and the code for their discovery process available to the public, inviting other scientists to use these tools to push the boundaries of what is possible. As the field of quantum computing moves forward, this kind of automated design may become the standard way to build the next generation of machines, turning the slow, manual art of circuit design into a fast, precise science.

Technical Summary: A General Framework for Gradient-Based Optimization of Superconducting Quantum Circuits

Problem Statement
Engineering the Hamiltonian of a quantum system is fundamental to achieving novel performance goals or realizing new functionalities in quantum computing. While superconducting quantum circuits offer substantial flexibility for Hamiltonian design, the manual design process is increasingly difficult as circuits grow larger and more intricate. Existing software packages facilitate analysis but do not inherently automate the design process. Furthermore, optimizing these systems requires computing gradients of eigenvalues and eigenvectors of large, sparse Hamiltonian matrices with respect to system properties (e.g., capacitance, inductance, Josephson junction energy). Calculating these gradients for arbitrary systems poses a significant computational challenge, particularly because standard automatic differentiation tools often lack support for the gradients of sparse matrix eigensystems. Consequently, design optimization often relies on zero-order methods (e.g., evolutionary strategies, Bayesian optimization) which are computationally expensive and lack the efficiency of gradient-based approaches.

Methodology
The authors present a comprehensive framework for gradient-based optimization of superconducting quantum circuits, leveraging and extending the open-source SQcircuit Python package. The methodology is built upon three core technical pillars:

Integration of Automatic Differentiation (AD): The authors integrated the PyTorch engine into SQcircuit. To overcome the limitation that standard AD libraries do not support gradients for sparse matrix eigenvalues and eigenvectors, they introduced a custom computational node within the PyTorch computational graph. This node utilizes first-order perturbation theory to analytically compute the derivatives of eigenvalues ( $f_i$ ) and eigenvectors ( $|f_i\rangle$ ) with respect to circuit parameters ( $x$ ):
$\frac{\partial f_i}{\partial x} = \langle f_i | \frac{\partial \hat{H}}{\partial x} | f_i \rangle$
$\frac{\partial |f_i\rangle}{\partial x} = \sum_{m \neq i} \frac{\langle f_m | \frac{\partial \hat{H}}{\partial x} | f_i \rangle}{f_i - f_m} |f_m\rangle$
This allows the framework to compute gradients of arbitrary loss functions constructed from circuit elements and the circuit's eigensystem.
Robust Optimization Pipeline: The authors developed a pipeline using the BFGS (quasi-Newton) optimization algorithm with random restarts to navigate the non-convex loss landscape. Key features of this pipeline include:
- Automatic Truncation Assignment: An algorithm that dynamically assigns Hilbert space truncation numbers for harmonic and charge modes based on eigenvector decay rates and ground state width estimates. This ensures computational resources are allocated efficiently to modes requiring higher resolution.
- Convergence Testing: A rigorous test to verify that eigenvalues and eigenvectors have stabilized with respect to the chosen truncation numbers before proceeding with gradient calculations.
- Reparameterization: Circuit element values are optimized in a logarithmic, unitless space to handle the wide dynamic range of physical parameters (e.g., femtofarads to picofarads) and prevent ill-conditioning of the loss landscape.
Qubit Discovery as a Case Study: The framework was applied to the discovery of single-qubit designs. The search was formulated as a constrained optimization problem maximizing the number of executable single-qubit gates ( $N = T \times G$ ), where $T$ is the decoherence time and $G$ is the upper bound on gate speed. Constraints were imposed on flux sensitivity ( $S_\phi$ ), charge sensitivity ( $S_n$ ), and element sensitivity to fabrication errors ( $S_e$ ) using hinge loss functions to penalize violations.

Key Contributions
The paper outlines four primary contributions:

Extended SQcircuit: The package was enhanced with automatic differentiation capabilities to compute gradients of arbitrary metrics with respect to circuit element values, including the novel implementation of eigenvalue/eigenvector gradients for sparse Hamiltonians.
Optimization Pipeline: A robust pipeline for circuit discovery using BFGS optimization, requiring only the definition of a suitable loss function.
Automatic Truncation and Convergence: Implementation of an automatic assignment of Hilbert space truncation numbers and diagonalization convergence tests, addressing significant bottlenecks in search processes.
Performance Metrics: Introduction of specific metrics to compare qubit designs based on laboratory implementability (sensitivities) and total gate count, applied to exhaustively explore single-loop circuit topologies.

Results
The framework was used to discover qubit designs that outperform existing literature counterparts in specific metrics:

Gate Count Improvements: Optimized circuits showed improvements in the heuristic measure of gate count ( $N$ ) compared to existing qubits. For example, optimized "JL" (fluxonium-like) circuits demonstrated gate counts an order of magnitude higher than heavy fluxonium and Pop's fluxonium designs.
Reduced Sensitivity: The optimized "JL" circuits exhibited significantly lower sensitivity to external flux variations compared to heavy fluxonium, mitigating a primary source of dephasing.
Topology Exploration: The pipeline explored various topologies (e.g., "JJ", "JL", "JJJ", "JLJL"). While some optimized designs (like "JLL" or "JL(JC)") initially appeared promising, the optimization process often reduced them to known, simpler effective circuits (e.g., shorting inductors to revert to fluxonium). However, the "JJJ" topology yielded a design with a gate count an order of magnitude higher than the flux qubit and quantronium, featuring two-dimensional transmon-like eigenvectors localized in the potential minimum.
Comparison: The discovered qubits were benchmarked against established designs (e.g., Sycamore transmon, heavy fluxonium, flux qubit) using a unified metric of gate count, decoherence time, and gate speed upper bounds.

Significance and Claims
The authors claim that this work provides a universal approach applicable to arbitrary superconducting quantum circuits, moving beyond problem-specific optimization strategies that require starting from scratch. By demonstrating that gradient-based optimization can be efficiently implemented for these systems, the framework streamlines the design process.

The paper modestly frames its significance as a demonstration of a general methodology. While showcased through qubit optimization, the authors state the framework is versatile and can be extended to other challenges in superconducting quantum hardware design, such as coupler design, quantum sensors, and transducers. The work does not claim to have solved all optimization challenges in quantum hardware but rather provides a scalable, automated toolset that accelerates the exploration of the design space. The authors note limitations, including the restriction to circuits with up to four nodes due to computational memory constraints and the simplification of gate speed bounds by neglecting coupling strengths and specific control constraints.

Technical Summary: A General Framework for Gradient-Based Optimization of Superconducting Quantum Circuits

More like this