Can gradient-based optimization automate superconducting qubit design and discovery?

Gradients of a circuit's eigensystem, not just its energy

The obstacle the paper targets is technical. A superconducting circuit's Hamiltonian is a large, sparse matrix whose entries depend on physical elements such as junction energies, capacitances, inductances, and external flux. Designing a qubit means choosing those element values so that derived quantities — transition frequencies, matrix elements, decoherence times — come out well. Gradient-based optimization needs the derivative of those quantities with respect to each element, and for eigenvectors of a large sparse matrix that derivative is not something a standard autodiff library supplies for free.

The framework's answer is to implement the eigensystem derivative as a custom computational node inside SQcircuit and switch the package's engine from NumPy to PyTorch. Once that is done, any circuit property expressed as a scalar metric can be differentiated by backpropagation. The paper's own tutorial example makes the operation concrete: a transmon is defined with a 12 fF capacitor flagged for gradient tracking and a 10 GHz junction left untracked, and after diagonalization the gradient of the matrix element g10 = (1/C)|⟨f1|Q̂|f0⟩| with respect to the capacitor is obtained by calling backward(). The resulting gradient carries units of the property per unit of the element — here, matrix element per farad.

Two implementation details matter for anyone reading the numbers later. Eigenvector gradient precision scales with the number of eigenstates retained, so truncation is not a neutral bookkeeping choice. And because first-order derivatives are custom nodes, PyTorch cannot chain them automatically into second derivatives; the authors add a separate node for the dephasing-time gradient, which depends on second derivatives of the Hamiltonian. That is why the framework can optimize decoherence time at all, rather than only frequencies.

Truncation allocation decides which circuits are trustworthy

Diagonalizing a circuit Hamiltonian requires truncating each mode's Hilbert space. The paper treats the total truncation number K as a proxy for Hilbert space size and shows that splitting K evenly across harmonic modes can fail: in one example with K = 3000, an even split of h1 = h2 = 14 left the slower-decaying mode unconverged, while a heuristic reallocation to h1 = 4 and h2 = 71 produced convergent flux spectra for the same budget.

The heuristic is validated by sampling 100 random circuits of several types and comparing the fraction that converge under even versus heuristic allocation, for "JLL" and "JLJL" topologies across a range of K. The heuristic yields a higher convergent proportion, reaching 100% convergence for a wide class of circuits. Charge modes were given the same even split under both schemes, so the comparison isolates the harmonic-mode allocation.

This is not a side detail. The pipeline's search over topologies depends on reliably diagonalizing whatever circuit it samples, and the paper sets each topology's K by initializing 100 random circuits and choosing the smallest truncation at which all 100 converge, with a maximum K of 1.6×10^4 needed for larger four-node circuits. A design that looks good only because its spectrum had not converged would be a false discovery, so the convergence test is part of the result, not an appendix formality.

The objective is a gate-count proxy with sensitivity penalties

The quantity being optimized is not a single physical measurement but a composite loss. The estimated number of gates N is the primary metric, and it depends on two other quantities: the decoherence time T and an upper bound on gate speed G, so that constant gate number corresponds to contours of GT. Qubits further toward high G and high T therefore score better on N.

Because N alone would reward designs that are fast and coherent but fragile, the loss adds penalties for sensitivity to external flux, charge, and circuit-element variation, with maximum tolerances set to S*φ = 0.1, S*n = 0.02, and S*e = 0.1 and loss weights βφ = βn = βe = βf = 1. Optimization runs 100 BFGS steps per circuit instance on Stanford's Sherlock cluster.

Element values are not optimized directly. Because capacitances can span picofarads to femtofarads, the authors reparameterize each element x through a log-scaled angle αx between 0 and 2π, keeping values inside fabrication bounds while making the landscape better conditioned. Readers comparing optimized element values to literature devices should keep this in mind: the reported numbers are the endpoints of a constrained search, not free-floating optima.

Rediscovering the transmon is the pipeline's validity check

The clearest evidence that the optimizer works as intended comes from the simplest topology. "JJ" circuits consist of two Josephson junctions forming an inductive loop, shunted by one capacitor, and are analytically tractable. The search found two such qubits. "JJ 1" sits at external flux near zero, with two nearly equal junctions (EJ1 = 1.15 GHz, EJ2 = 1.04 GHz) and a large 101 fF shunt capacitance — the paper reads this as a rediscovery of the transmon.

"JJ 2" is the more informative case. It operates at φ*ext = π, an unusual and normally flux-sensitive working point for a transmon, but the optimizer selected asymmetric junctions (EJ1 = 6.79 GHz, EJ2 = 1.00 GHz, C = 38.1 fF) that reduce flux sensitivity there. This is the pipeline doing design work rather than pattern matching: it found a configuration that satisfies the sensitivity penalty at a point where the standard device would not.

Both optimized "JJ" qubits show an order of magnitude larger gate count than the Sycamore transmon (EJ1 = EJ2 = 18.0 GHz, C = 108 fF). The authors explicitly frame this as a comparison of each topology's inherent potential, noting that Sycamore may have been optimized for other functionality such as two-qubit gates. That caveat should travel with the number: the gate-count gap is not a claim that a fabricated Sycamore-class device is outperformed.

Fluxonium-class circuits win on this objective

The "JL" topology — one junction, one superinductor, one capacitor — contains the fluxonium. Optimized "JL" circuits reach a gate number an order of magnitude above both Pop et al.'s fluxonium and the heavy fluxonium, and the paper predicts they are significantly less sensitive to external flux at their operating points. The element values are notably different from the literature devices: "JL 1" uses EJ = 6.86 GHz, L = 1.61 µH, C = 9.76 fF and "JL 2" uses EJ = 4.26 GHz, L = 1.83 µH, C = 8.08 fF, against heavy fluxonium's EJ = 3.39 GHz, L = 1.23 µH, C = 40.4 fF.

The flux-sensitivity claim is the substantive one. The paper notes that heavy fluxonium's metrics imply a high gate count but that it suffers from flux sensitivity, which the loss penalizes; the optimized "JL" circuits avoid that penalty, and their flux spectra are plotted against heavy fluxonium to show reduced sensitivity at the operational flux point. The authors describe the resulting element range as not previously investigated.

The paper's own summary of the search is that fluxonium has the best overall properties for the objective function considered. That is a statement about this loss, these element bounds, and these topologies — not a general ranking of qubit architectures.

Where the optimization stops short of a real device

The gate-speed bound G is built from frequency spacing between states and does not include coupling strength between them, which the authors identify as critical when drive power is limited. It also omits control constraints and leakage induced by the drive source. So a high G in this framework means the levels are favorably spaced, not that a working gate has been demonstrated.

The search space itself is bounded in three ways that shape every result above. Circuits are limited to four nodes, because memory requirements grow exponentially with node count; the authors point to tensor-network representations as a possible remedy. The logical state is assumed to be encoded in the ground and first excited states, excluding bosonic or dual-rail encodings. And element values are constrained to fabrication-informed ranges.

The "JJJ" results illustrate a further practical limit: optimizing 100 instances of the three-junction topology produced final losses sensitive to the initial sampled parameters, with a significant majority converging to suboptimal values and the best circuit shown as a solid red line among dashed suboptimal ones. Gradient descent here finds good local optima, not guaranteed global ones — consistent with the paper's own framing that non-convex loss functions make global optima difficult.

Read the gate-count gains as topology potential

The paper's most decision-relevant claim is comparative and conditional: for a loss that weights estimated gate count against flux, charge, and element sensitivity, fluxonium-class "JL" circuits came out ahead, and the optimizer found element values outside the ranges used by existing fluxonium and heavy-fluxonium devices. The supporting mechanism is concrete — reduced flux sensitivity at the operating point, achieved through junction asymmetry and reparameterized element search rather than through a new physical ingredient.

The right way to use the order-of-magnitude gate-count comparisons is as a statement about what each topology can support under this objective, exactly as the authors frame the Sycamore comparison. Whether an optimized "JL" design retains its advantage once coupling strength, drive-induced leakage, and readout are modeled is left open by the paper's own limitations, and no fabricated device is reported.

What the work does establish is that the gradient machinery is general: the eigensystem derivative, the custom second-derivative node for dephasing, and the convergence-aware truncation scheme are tied to SQcircuit's circuit representation rather than to qubits specifically, and the authors propose applying the same pipeline to couplers, readout, gate implementations, sensors, and transducers. The transferable result is the optimization loop; the qubit rankings are its first demonstration.

Sources used in this answer