GAnnET: A principled approach to automatically annotating charge stability diagrams
Authors: Brian Weber, Justyna P. Zwolak
Organizations: Intelligent Geometries, LLC, Lake Frederick, VA 22630, USA · National Institute of Standards and Technology, Gaithersburg, MD 20899, USA · Joint Center for Quantum Information and Computer Science, University of Maryland, College Park, MD 20742, USA
Gate-defined semiconductor quantum-dot (QD) arrays are a promising platform for quantum computing. However, their large configuration spaces and inherent measurement noise make device tuning time consuming and difficult to scale. Although machine-learning (ML) approaches for automated tuning have proliferated rapidly, a fair comparison of methods remains limited by inconsistent metrics and a lack of standardized labeled datasets. Here, we introduce Geometry-Informed ANNotator for Electrostatic Topologies (GAnnET), a physics- and geometry-informed algorithm that automatically generates probabilistic state labels for charge stability diagrams (CSDs) of double-QD systems. GAnnET constructs local geometric models of CSD structure and combines them into a global state map with calibrated confidence, eliminating the need for manual annotation. We validate GAnnET on simulated and experimental CSDs and demonstrate consistent labeling suitable for ML training and benchmarking. By enabling reproducible annotation of experimental data, GAnnET supports community benchmarking and scalable automated QD operation.
Figures & tables
Figure 1: Schematic CSD and corresponding global state map. (A) Schematic CSD showing the irregular polygonal tiling of the (VP1,VP2) configuration space. The (m,n) labels denote the ground-state charge configuration of the left and right QDs. The inset shows a schematic potential landscape in the DD state. (B) Corresponding global state map partitioning of the (VP1,VP2) configuration space into five domains: ND, SD L , SD C , SD R , and DD.
Figure 2: Preprocessing of the binary transition mask for a sample simulated (top) and experimental (bottom) CSD. (A) Simulated 2D measurement generated with the QD simulator [ 41 ] . The inset shows the corresponding ground-truth state labels from the simulation. (B) Numerical derivative of (A), highlighting charge-transition features and revealing the irregular polytope tiling. (C) Binary transition mask derived from (A), which serves as the required GAnnET input. (D) Experimental 2D measurement from the QFlow 2.0 dataset [ 32 ] . (E) Numerical derivative and (F) binary transition mask derived from (D).
Figure 3: The polygon modeling. (A) A preview of a fingerprint measurement from a sample point (shown in pink) from the initial grid of observation points. A preview of the initial grid of sampled points is shown in gray. (B) A series of consecutive observation points selected by the model-building module, with the intensity of the color indicating the progress of the algorithm. (C) Model creation for a nearly ideal polygon (right) and the resulting polygon (left). (D) A corrected set of rays for a noisy model of a polygon (right) and the resulting polygon model (left).
Figure 4: The depiction of the dominant directions in CSD. (A) Fingerprint-modeled polygons for the CSD shown in Figure 2 D, showing overlaps and gaps, and some misidentifications caused by noise and boundary effects. Overlaps are resolved by assigning each contested point to the observation point closest to it. Gaps are resolved by the expansion method described in the text. (B) The depiction of the characteristic left, center, and right dominant directions of the (VP1,VP2) configuration space. (C) The “filled-in” map. The red, green, and blue colors indicate left, central, and right orientation, respectively. The gray color in (A) and (C) indicates the ND domain, while the black color indicates pixels that were not assigned to any polygon during the modeling phase.
Figure 5: The performance of GA NN ET on simulated CSDs with varied noise levels. The noise levels used in all tests with simulated data range from 0.00 to 5.00 , where the reference noise level extracted from the experimental data is denoted, for simplicity, as 1.00 . The boxes span the interquartile range ([IQR], 25th–75th percentiles), the whiskers extend to 1.5 times the IQR from the box limits, and the horizontal lines inside boxes denote the median. Points outside the whiskers represent outliers. The inset shows the overall performance across all domains and CSDs at each noise level. Data are presented as means, and error bars represent the sample standard deviation.
Figure 6: Final probabilistic domain decomposition for simulated and experimental CSD. (A) A sample noisy simulated CSD shown in Figure 2 A and (C) a sample experimental scan shown in Figure 2 D overlaying the charge sensor data. Regions with confidence in the dominant label of at least 70% are represented as ringed areas for (B) the simulated data and (D) the experimental scan.
Figure 7: Two kernel choices in the heat flow clustering method. The x axis parameterizes S1 , and the points are polygon orientations. The inset in each plot shows the kernel choice. Properly selected kernels can lead to clear, persistent clustering of points, with each point ultimately assigned to the cluster in which it appears most persistently. L , C , and R indicate left, center, and right orientation clusters, respectively.
Figure 8: Measuring the hexagonness of cells. (A) The division of the hexagonal cell into an upper (above the centerline) and lower (below the centerline) polygon showing the cell roof (pink) and floor (green). (B) The polygonal cell’s dimensions in terms of physical quantities in the constant-interaction model. [ 36 ]
Jul 23, 2026·Hyma Vallabhapurapu, Marco Candido, Krishna Choudhary +7
Diraq, Sydney, New South Wales, Australia · School of Electrical Engineering and Telecommunications, University of New South Wales, Sydney, New South Wales, Australia