Skip to main content
Back to timeline
arXivSource publication:

QCap turns superconducting device capacitance extraction from hours into seconds, holding a 0.177% median error on 4,009 held-out designs

Synopsis

Extracting the capacitance matrix of a superconducting quantum device normally means running an electromagnetic field solver on every layout, and a single qubit with several couplers can take hours. QCap converts a layout into a graph whose nodes are conductors and whose edges are conductor pairs, obtains a coarse physics prior from a fast method-of-moments solve, and trains a graph neural network to learn only a bounded residual correction. On 4,009 designs withheld from training, its output agrees with Ansys Q3D at a median relative error of 0.177%, and on entirely unseen component families the error stays below 1%.

Source-provided article image: QCap: a physics-corrected graph neural network for the capacitance matrix of any planar superconducting device layout
arXiv

Interpretation

QCap produces the full Maxwell capacitance matrix directly from the .gds layout of any planar superconducting device, without retraining per component family and without depending on the drawing tool. Earlier machine-learned surrogates took the parameter string of one layout-generation tool as input, so they could only be evaluated on the component they were trained on, and each new component required thousands of new simulations. On 4,009 test designs withheld from training, QCap agrees with Q3D at a median relative error of 0.177%; on component families never seen in training the error is 0.33% to 0.99%.

The model encodes a layout as a complete graph whose nodes are conductors, ground included, and whose edges are conductor pairs, and the same functions act on every node and edge, so the parameter count does not depend on the number of conductors. A single fixed-length vector cannot carry a layout whose conductor count varies, and a parameter-string representation is tied to a particular tool. The conductor vector holds shape scalars, Fourier magnitudes of the outer contour and a width spectrum; the edge vector records a separation spectrum of facing boundary, coupling moments, a coplanar coupling and shielding shares; in a 67-conductor grid, the same qubit alone on its chip and inside the grid has every entry identical except the two shielding shares.

The network learns only a bounded logarithmic residual between the physics prior and the reference solution, where the prior comes from a fast method of moments on the metal and one Richardson step suppresses its mesh error. Learning Maxwell's equations directly from data is a hard training problem that demands large amounts of simulation data. The prior takes a median of 0.075 s on two CPUs, against 5.2 to 52 minutes per layout for Q3D; on validation layouts the Richardson step halves the prior's error at less than twice the cost of the finer solve alone.

For layouts with more than nine conductors the prior is solved in windows, and distant conductor pairs outside any shared window are filled with a closed-form grounded-plane dipole formula, so the prior cost grows with conductor count rather than chip area. A dense method-of-moments solve does not scale to large chips, and no training layout has more than 68 conductors. On the 67-conductor grid, the windowed result differs from a global solve on the same mesh by 0.3% to 0.6% between signal conductors; the nine-qubit, 67-conductor device yields its matrix in 11.7 s on a laptop, against 19.9 h and 228 GB of memory in Q3D.

Perspective

The demonstrated scope is quasi-static capacitance extraction for planar superconducting devices, with all training data on a 350-micrometer silicon substrate and no layout above 68 conductors; the prediction for the 100-qubit, 781-conductor chip has no reference solution, so it is not an absolute-accuracy claim.

The 100-qubit chip result has no Q3D reference, and the authors explicitly make no absolute-accuracy claim there; error on far conductor pairs is limited by the prior, so the network applies its maximum correction; training data cover a single substrate, and other substrates rely on substrate information in the prior plus a small correction factor.