The Signal
Stanford Team Uses Automatic Differentiation to Search for Qubits
A Stanford-led team published a gradient-based framework for optimizing superconducting circuits in SQcircuit. Its simulations identify fluxonium-like designs with improved estimated gate counts, but the proposed devices still need laborato
Quantum.world Team · Written with AI · fact-checked ·
A Stanford-led team reports simulated superconducting-qubit designs with an estimated gate-count metric an order of magnitude above selected literature counterparts. But the work does not report fabricated devices or laboratory measurements: its performance claims come from numerical optimization, and its gate-speed estimate leaves out factors needed to predict practical operation.
The study, published in Quantum on October 27, 2026, adds automatic differentiation to the open-source SQcircuit software package. The authors use the framework to search for circuits that balance coherence and gate speed against sensitivity to noise and fabrication variation. Their own account points to the gap between design and experiment: “We have oriented our approach to develop experimentally realistic designs, and hope to use this strategy to create better devices for future laboratory tests.” Taha Rajabzadeh et al., “A General Framework for Gradient-Based Optimization of Superconducting Quantum Circuits using Qubit Discovery as a Case Study”
Gradients for sparse circuit Hamiltonians
Superconducting circuits can be tuned by changing component values and topology, but finding a design that satisfies several performance and fabrication constraints means searching a large parameter space. The researchers’ method calculates how a circuit’s properties change when those parameters change, then uses the resulting gradients to guide the search.
The work extends SQcircuit, a Python package for analyzing superconducting circuits, with a custom calculation step that works with PyTorch’s automatic-differentiation engine. The added step addresses a specific software obstacle: PyTorch does not directly provide gradients for the eigenvalues and eigenvectors of the sparse Hamiltonians used in SQcircuit. The framework computes those derivatives separately and feeds them into the optimization process.
For the qubit-discovery demonstration, the team uses BFGS, a gradient-based optimization method, with random starting points and parallel runs across CPU cores. The search includes convergence checks and automatically assigned Hilbert-space truncations, intended to keep the circuit simulations accurate enough for comparison. The authors say the framework can also optimize other circuit properties by changing the objective function.
Fluxonium leads within a bounded search
The discovery study examines single-loop circuits with up to four nodes, rather than every possible superconducting-circuit topology. It constrains component values to ranges the authors treat as fabricable and compares candidate designs using estimated gate count, coherence time, an upper bound on gate speed, and sensitivity to noise and component variation.
Among the circuits in that search, the authors find that fluxonium-type designs—built around a Josephson junction and a superinductor—score well under their chosen objective. They report that optimized designs have an estimated gate-count metric about an order of magnitude above selected literature examples. The comparison is not a claim that the new devices outperform those qubits in every use: the paper says existing designs may have been optimized for other functions, including two-qubit gates.
The score has limits—and the hardware remains untested
The gate-count metric combines coherence time with the estimated upper bound on gate speed. The authors’ speed estimate relies on energy-level spacing and does not account for coupling strength, control constraints, or some sources of leakage from the drive. They note that including more detailed models of control and readout could make the predictions more precise for practical designs.
The search also covers only circuits with up to four nodes, and the authors say larger simulations exceeded the computational resources available to them. Its results therefore identify candidates within defined boundaries, not a universal ranking of qubit architectures. The next concrete test is whether a design selected by the software can be fabricated and measured; the paper presents that as future laboratory work, not a result already achieved.
The revised preprint appeared on arXiv on October 13, 2026, before the journal publication two weeks later. SQcircuit and the authors’ qubit-discovery code are available online, giving other researchers a way to inspect or adapt the optimization pipeline.