Speaker
Description
Tuning semiconductor spin-qubit arrays becomes intractable as devices scale from a few quantum dots to arrays of tens: cross-talk between gates forces manual tuning that already takes days for an 8-qubit register. Existing automation analyses each charge stability diagram (CSD) in isolation under the constant-interaction model, which fixes capacitances and discards tunnel coupling, the parameter essential for two-qubit gates.
We instead frame multi-dot tuning as a set-learning problem. Each CSD is a 2D cross-section of the polytopes of constant charge, living in higher-dimensional gate-voltage space. Tuning means reconstructing that structure from a limited set of slices. We train a Set Transformer that jointly processes an arbitrary number of CSDs, learning correlations across them and pooling into a shared latent representation. From this single latent, a multi-task head simultaneously predicts lever arms $\alpha$, charging energies $E_c$, and tunnel couplings $t_c$, to our knowledge the first architecture to extract quantum corrections directly from CSDs.
On simulated four-dot arrays, average nearest-neighbour tunnel-coupling $R^2$ rises from $0.81$ (4 cuts) to $0.90$ (6 cuts), evidence that the model fuses complementary projections rather than reading each in isolation. The architecture is permutation-invariant and variable-size, requiring no retraining as the number of measurements grows.
| Topic | Machine learning techniques for experimental quantum information science |
|---|