SimCATS: Simulator for Quantum Dot Tuning

Updated 4 July 2026

SimCATS is an open-source Python framework that generates realistic charge stability diagrams of quantum-dot devices using a modular, measurement-oriented approach.
It employs a geometric occupation model combined with a sensor-response and distortion modules to replicate key features observed in experimental data.
The framework supports automated tuning, benchmarking, and robustness studies by providing fast, labeled synthetic datasets for quantum computing research.

SimCATS, short for “Simulation of Charge Stability Diagrams for Automated Tuning Solutions,” is an open-source Python framework for generating realistic simulated charge stability diagrams (CSDs) of quantum-dot devices, especially double quantum dots (DQDs), together with ground-truth charge information. Its stated purpose is to support automated tuning of quantum dots by supplying abundant, configurable, labeled data for algorithm development, testing, benchmarking, and robustness studies. The framework is explicitly measurement-oriented rather than ab initio: it generates an idealized charge-occupation structure, computes a sensor response from that occupation, and then applies experimentally motivated distortions and noise, with code released at https://github.com/f-hader/SimCATS (Hader et al., 11 Aug 2025).

1. Position within automated quantum-dot tuning

The motivation for SimCATS is tied to the scaling problem in semiconductor quantum computing. Quantum dots must be tuned very precisely before they can be used as qubits, and two stages are emphasized: dot-regime tuning, where the device is brought into the desired few-dot configuration, and charge-state tuning, where the correct electron occupation is established. In experiments, this is usually done by analyzing CSDs obtained from a nearby charge sensor. Manual tuning does not scale to large arrays, so automated tuning algorithms—classical and machine-learning based—require large and diverse datasets with reliable labels. Experimental datasets are expensive to collect and often require manual labeling; simulated datasets can fill that gap if they are realistic enough (Hader et al., 11 Aug 2025).

SimCATS is designed specifically for that intermediate regime. It does not try to be a full quantum transport solver, and it is not framed as a tool for physical parameter extraction. Instead, it aims to reproduce the shapes and artifacts actually seen in measured DQD CSDs while remaining much faster and easier to parameterize than detailed physical simulators. This design choice makes the framework sample-agnostic and comparatively easy to fit from existing data. A plausible implication is that SimCATS is intended less as a device-theory platform than as an infrastructure layer for autotuning research, where realism of the measurement image and availability of labels are often more important than microscopic faithfulness.

The framework is also intentionally broad in measurement configuration. It supports 2D and 1D measurements, different scan directions, time-dependent distortions, different sensor configurations including multi-sensor setups, and the option to enable or disable individual distortion mechanisms. Ground-truth outputs include ideal occupation data in addition to simulated sensor-response images. The paper explicitly positions this as useful for training and testing machine-learning models for charge-state detection, developing algorithms that identify the correct electron number in dots, benchmarking published tuning algorithms on common synthetic datasets, studying robustness to sensor sensitivity and varying distortion strengths, evaluating data quality control and device-state estimation pipelines, and supporting active measurement strategies or virtual-gate extraction workflows (Hader et al., 11 Aug 2025).

2. Measurement-oriented simulation architecture

A defining feature of SimCATS is its modular workflow. Experimental CSDs are treated as the result of three conceptually distinct stages: an occupation model, a sensor-response model, and a distortion model. Each stage is implemented through simple interfaces, so standard components can be replaced by custom models. This division is central to the framework’s extensibility and to its claim of technology independence (Hader et al., 11 Aug 2025).

The first stage constructs an ideal occupation map. The second converts that occupation into a simulated sensing-dot response, incorporating both occupation-induced signal changes and gate cross-coupling. The third injects realistic distortions and noise so that the final output resembles measured data rather than a clean synthetic diagram. The framework can therefore simulate not only the “physics-like” honeycomb pattern but also the measurement artifacts that often dominate autotuning performance in practice.

This decomposition also structures how realism is controlled. Users can keep the occupation map fixed while varying sensor sensitivity, or keep the sensor model fixed while sweeping distortion strength. Because several distortion types are time dependent, scan direction matters: the same underlying charge structure can produce different artifact patterns when line-wise measurement order changes. SimCATS supports different scan directions explicitly for that reason.

A further practical element is the inclusion of extracted parameter sets. The paper states that a default parameter configuration derived from a GaAs sample is bundled as default_configs["GaAs_v1"]. This indicates that the framework is not only a code library but also a repository of empirically motivated parameterizations that can serve as starting points for reproducible benchmarking.

3. Geometric occupation model for double quantum dots

The occupation model is the most distinctive part of SimCATS. Rather than deriving the honeycomb structure from a constant-interaction or Hubbard Hamiltonian with explicit capacitances, the framework introduces a purely geometric model. The goal is to reproduce the shapes seen in measured DQD CSDs using a compact set of geometric parameters, without requiring detailed knowledge of physical device parameters (Hader et al., 11 Aug 2025).

The central objects are total charge transitions $tct_i$ , where $tct_i$ separates regions containing $i-1$ and $i$ total electrons. In a DQD CSD, these transitions are built from lead-to-dot transitions (LDTs) and curved regions around triple points caused by interdot tunnel coupling. To describe slopes robustly, the paper rotates the original plunger-gate voltage space $(V_{P1},V_{P2})$ by $45^\circ$ to obtain $(V'_{P1},V'_{P2})$ . In that transformed space, one LDT family has slope in $[-1,0)$ and the other in $(0,1]$ , which makes the two families easy to distinguish.

Each $tct_i$ is parameterized by four classes of geometric quantities: the $tct_i$ 0-intercepts $tct_i$ 1 of the two LDTs, their slopes $tct_i$ 2, and two Bézier anchor points $tct_i$ 3 and $tct_i$ 4 that define the beginning and end of the curved part around the first triple point. A total charge transition is then assembled by repeating linear pieces, a Bézier curve, and its $tct_i$ 5-rotated counterpart, yielding a curve that is twice continuously differentiable around triple points. The maximum number of triple points or Bézier-curve segments on $tct_i$ 6 is limited by

$tct_i$ 7

Once the family of total charge transitions has been defined, the ideal occupation map is obtained geometrically. The region between $tct_i$ 8 and $tct_i$ 9 contains $i-1$ 0 total electrons. Their partition between the two dots is inferred from interdot transitions $i-1$ 1, each defined by the line connecting a triple point of $i-1$ 2 to the opposite triple point of $i-1$ 3. Across each such line, a sigmoid function orthogonal to that transition approximates a Fermi-like crossover of charge between the dots. The superposition of these sigmoids yields the occupation $i-1$ 4 of the first dot, and the second-dot occupation follows from total charge conservation:

$i-1$ 5

The paper is explicit that this model is approximate. It reconstructs measured shape rather than solving electrostatics or quantum transport. It is therefore especially suitable when the objective is realistic labeled data generation for tuning tasks, not direct inference of capacitances, tunnel couplings, or Hamiltonian parameters. This suggests that SimCATS deliberately privileges controllable morphology and dataset scale over microscopic interpretability.

4. Sensor response and distortion model

Experimental CSDs are not direct plots of occupation; they are measurements of a nearby sensing dot or sensor conductance/current. SimCATS models this through a compact sensor-response stage in which the sensor electrochemical potential is shifted both by quantum-dot occupation and by direct gate cross-coupling:

$i-1$ 6

followed by a simplified Lorentzian Coulomb peak:

$i-1$ 7

Here $i-1$ 8 is the occupation of dot $i-1$ 9, $i$ 0 is the plunger-gate voltage, $i$ 1 is the lever arm from dot occupation into the sensor potential, $i$ 2 is the gate cross-coupling into the sensor potential, $i$ 3 is the baseline sensor electrochemical potential, $i$ 4 is the simulated sensor signal, $i$ 5 is an output offset, $i$ 6 is a scaling factor, $i$ 7 is the Lorentzian width, and $i$ 8 is the electrochemical potential at the Lorentzian peak (Hader et al., 11 Aug 2025).

The paper states that $i$ 9 mainly controls edge sharpness, $(V_{P1},V_{P2})$ 0 mainly controls in-cell drift, and these influences are typically counteractive in sign, with $(V_{P1},V_{P2})$ 1 and $(V_{P1},V_{P2})$ 2. This allows the same occupation geometry to generate sensor images with visibly different edge contrast and interior cell gradients.

Distortions are grouped according to where they enter the simulated signal path:

Distortion class	Examples	Injection stage
Occupation distortions	Dot jumps; occupation-transition blurring	Occupation map
Sensor-potential distortions	Pink noise; RTN	Sensor potential
Sensor-response distortions	White noise; category-3 RTN	Final sensor signal

Occupation distortions include dot jumps and occupation-transition blurring. Dot jumps are modeled as charge-trapping events that shift blocks of columns horizontally or blocks of rows vertically; their extension is sampled from a geometric distribution and their amplitude from a Poisson distribution. Occupation-transition blurring represents thermal broadening of lead transitions due to the Fermi-Dirac distribution in the reservoir, but the implementation reported in the paper approximates the intended Fermi-Dirac-like filter by a Gaussian kernel.

Sensor-potential distortions include pink noise and random telegraph noise (RTN). Pink noise is generated with a standard power-law-noise method and becomes more visible in high-gradient regions because it is injected before the nonlinear Lorentzian response. RTN is modeled as bursts switching between discrete levels; burst extension is sampled geometrically and amplitude from a normal distribution. The paper notes that RTN can appear either in sensor potential or directly in sensor response, and SimCATS supports both variants.

Sensor-response distortions include white noise and category-3 RTN. White noise represents thermal and shot noise after amplification and is modeled as Gaussian with standard deviation $(V_{P1},V_{P2})$ 3. For PSD-based parameterization, the paper gives

$(V_{P1},V_{P2})$ 4

and fits high-frequency white and pink contributions jointly through

$(V_{P1},V_{P2})$ 5

with $(V_{P1},V_{P2})$ 6 and $(V_{P1},V_{P2})$ 7 for Welch’s method as used through SciPy. Pink noise is thus modeled as $(V_{P1},V_{P2})$ 8-like in PSD, while RTN is described qualitatively as having $(V_{P1},V_{P2})$ 9-like PSD.

5. Parameter extraction, validation, and benchmarking

SimCATS includes a workflow for extracting model parameters from measured data. For occupation geometry, total charge transition parameters can be manually labeled from recorded CSDs, and transformation rules can then be defined to generate successive transitions rather than storing them individually. For the sensor model, Lorentzian parameters are fitted from measured sensor scans. Assuming operation on the monotonic flank of the Lorentzian, the sensor potential can be estimated from measured sensor values, after which cross-coupling coefficients and occupation lever arms can be extracted inside fixed-charge and lead-transition regions, respectively (Hader et al., 11 Aug 2025).

The framework is validated both qualitatively and quantitatively. The visual validation shows sequences of neighboring 2D simulated scans during an iterative tuning procedure, with honeycomb structures evolving across voltage space in a way that qualitatively resembles experiment. For quantitative validation, the paper borrows generative-model metrics: $45^\circ$ 0-precision and $45^\circ$ 1-recall. Because standard image-generation metrics such as Inception Score or FID were deemed unsuitable without a domain-specific pretrained classifier, the authors trained a custom feature embedding network based on an adapted MNIST_LeNet architecture and Deep SVDD, using the loss

$45^\circ$ 2

Training used 50% of 484 experimental CSDs, with the remaining 50% used for testing, and the training data were augmented by rotations, flips, and brightness/contrast changes. As a sanity check, experimental training versus test achieved high precision and 100% recall; MNIST data mapped into the same feature space gave high precision but very low recall. For SimCATS, the reported table gives $45^\circ$ 3-precision of $45^\circ$ 4 and $45^\circ$ 5-recall of $45^\circ$ 6, while an expanded SimCATS dataset improves recall to $45^\circ$ 7. The interpretation given is that the simulator captures a large fraction of the measured distribution but not all experimental variations.

The paper also benchmarks runtime and memory against two physical simulation approaches implemented in QuDiPy: a constant-interaction model and a Hubbard model. For resolutions from 50 to 500 pixels per axis, SimCATS required about 19–418 ms and 44.4–90.0 MB memory. After initialization, the physical-model runtimes increase roughly quadratically with image resolution, whereas SimCATS remains much faster and depends only weakly on resolution because total-charge-transition construction dominates and per-pixel evaluation is inexpensive. This is one of the framework’s main practical advantages for large-scale dataset generation and iterative algorithm development.

6. Scope, limitations, and nomenclatural distinction

The paper states several assumptions and limitations explicitly. SimCATS assumes reasonably good samples suitable for scalable devices: no severe parasitic dots under barriers, no strongly moving dots, and sufficiently slow measurements so that latching can be neglected. The occupation model is geometric rather than microscopic, so it does not directly infer capacitances, tunnel couplings, or Hamiltonian parameters. Thermal broadening is conceptually Fermi-Dirac but was approximated by a Gaussian in the reported implementation. Dot-jump and RTN parameter extraction was largely manual because robust automatic detection was not yet available. Broader effects such as varying transition-line lengths over very wide scans, inter-scan correlations, or simultaneous multi-DQD correlations are not central in the present version, although the modular design could accommodate them. The relation between geometric total-charge-transition parameters and parameters of physical models is identified as future work (Hader et al., 11 Aug 2025).

These limitations are inseparable from the framework’s intended role. SimCATS is presented as filling a gap between highly realistic but slow physical simulators and simplistic synthetic datasets that omit crucial measurement artifacts. Its usefulness therefore derives from a deliberate abstraction: geometric charge-structure generation, compact sensor modeling, and modular distortion injection.

A recurrent source of confusion is acronym overlap. SimCATS is distinct from several unrelated systems that use the acronym CATS on arXiv and do not introduce SimCATS by that name, including the Cybersecurity Assessment Tools project (Sherman et al., 2017), Cost-Augmented Monte Carlo Tree Search for LLM-assisted planning (Zhang et al., 20 May 2025), CATS for distributed transformer inference on ultra-low-power wireless devices (Gräfe et al., 15 May 2026), Conductor-driven Asymmetric Transport Scheme in transport systems (Rizvi, 14 Mar 2026), its later mixed-criticality transport extension (Rizvi, 15 Jun 2026), and Converged Accelerated Taylor Series for nuclear reactor point kinetics (Ganapol, 2024). In the quantum-dot context, by contrast, SimCATS refers specifically to a simulator of charge stability diagrams for automated tuning solutions.

Taken together, these features define SimCATS as a fast, modular, open, measurement-oriented simulator of CSDs that includes ideal charge structure, sensing-dot readout, and experimentally motivated distortions. Its central contribution is not a microscopic theory of quantum-dot devices, but a configurable synthetic-data framework for autotuning research, benchmarking, and robustness analysis.