SimCATS: Synthetic Quantum Dot CSD Simulator

• SimCATS Framework is a Python-based simulator that generates realistic synthetic charge stability diagrams (CSDs) for semiconductor double quantum dots, aiding automated tuning and machine-learning control.
• It employs modular models for dot occupation, sensor response, and signal distortions, enabling rapid, parameterized generation of labeled datasets that mirror experimental conditions.
• Its extensibility supports benchmarking tuning algorithms and cross-domain applications, though its geometric approach limits detailed physical predictiveness.

Searching arXiv for papers mentioning "SimCATS" to ground the article and disambiguate usages. SimCATS most explicitly denotes a Python framework for simulating realistic charge stability diagrams (CSDs) of gate-defined semiconductor double quantum dots, introduced to support automated tuning and machine-learning based control in semiconductor qubit platforms (Hader et al., 11 Aug 2025). Its stated role is not that of a full quantum-transport solver, but of a fast, flexible generator of measurement-like data that combines an idealized occupation model with sensor response and realistic distortions. In adjacent literature and structured syntheses, the label “SimCATS” is also used more loosely for a microscopic traffic-simulation implementation of the CATS framework and as a design shorthand for a consistency–accuracy analysis toolkit inspired by CAT; this suggests that the term is polysemous across research domains rather than uniquely fixed to a single formalism (Wang et al., 2021, Cavalin et al., 26 Nov 2025).

1. Definition, purpose, and scope

In its explicit, titled usage, SimCATS addresses a bottleneck in scaling quantum-dot qubit platforms: the need to automate charge-state tuning using CSDs as the primary diagnostic representation (Hader et al., 11 Aug 2025). CSDs are described as 2D maps of sensor response or current versus two gate voltages, and they provide the patterns that human operators and tuning algorithms use to detect the presence of dots and their coupling, read off total electron number and its distribution, and extract quantities such as virtual gate transformations, lever arms, and tunnel couplings.

The framework is proposed as a generic mechanism for generating large numbers of synthetic CSDs together with ground-truth labels, including occupations and transitions. The stated motivation is methodological as much as practical: real experiments can provide limited and partly biased datasets, whereas automation methods require large labeled collections for supervised learning, robustness validation, and comparative benchmarking. SimCATS is therefore positioned as a shared tool for fair and reproducible evaluation of tuning algorithms, while remaining parameterizable by measured data from different platforms such as GaAs and Si (Hader et al., 11 Aug 2025).

This scope is deliberately narrower than a fully predictive microscopic device simulator. The framework is described as technology-independent in the sense that it does not explicitly encode GaAs versus Si/SiGe versus MOS physics, but instead relies on parameter sets fitted to measured CSDs and sensor behavior. A plausible implication is that its principal value lies in realistic dataset generation and stress testing of automation pipelines rather than in inferring device physics from first principles.

2. Architectural organization

The SimCATS “simulation class” is organized around three main components: a dot occupation model, a sensor model, and a distortions model (Hader et al., 11 Aug 2025). The occupation model computes the electron occupation of each dot across gate-voltage space; the sensor model maps occupations and gate voltages to a sensor electrochemical potential and then to an observable signal; the distortions model injects realistic perturbations at different stages of the signal path.

This architecture is intended to emulate a virtual measurement engine. A voltage sweep, either one-dimensional or two-dimensional, is specified; the occupation model is evaluated on that sweep; the sensor model converts occupations into a charge-sensor output; distortions are then applied according to configured parameters; and the result is a simulated CSD or line scan resembling experimental data. The framework supports ideal, noise-free outputs as well as controlled combinations of noise and artifacts, which is central to its use in benchmarking.

The design emphasizes extensibility through interfaces or plug-ins. Each of the three main models is defined through simple Python interfaces, and the package ships with default implementations while permitting user-supplied replacements that satisfy the same interface contract. The framework also supports 2D scans, 1D scans along arbitrary directions in gate space, and sequential scans with time-dependent distortions. The codebase is distributed openly, with a default GaAs parameter set provided as default_configs["GaAs_v1"] (Hader et al., 11 Aug 2025).

A concise summary of the core modules is given below.

Component	Role	Default characterization
Dot occupation model	Computes electron occupation over gate-voltage space	Geometric description of charge-transition lines
Sensor model	Converts occupation and voltages into sensor signal	Sensor potential plus Lorentzian response
Distortions model	Applies realistic perturbations	Occupation, sensor-potential, and sensor-response distortions

3. Geometric occupation model and charge-state representation

A defining design choice is that SimCATS does not use a capacitance matrix or a full constant-interaction or Hubbard model to compute energy levels (Hader et al., 11 Aug 2025). Instead, it adopts a geometric parametrization of the CSD based on experimentally observed total-charge transition lines. This yields a representation that is explicitly fast and configurable, while still reproducing the characteristic honeycomb structure of double quantum dots.

The framework organizes the CSD in terms of total charge transitions $tct_i$, where $i$ counts the total number of electrons in the combined double-dot system. Each $tct_i$ separates regions containing $i-1$ and $i$ electrons. In the 2D gate-voltage plane, a given $tct_i$ consists of two lead-to-dot segments, $ldt_{i,1}$ and $ldt_{i,2}$, together with one or more curved segments at triple points that capture tunnel-coupling effects. The geometry is defined in a rotated voltage space $(V'_{P1}, V'_{P2})$, obtained from the original plunger-gate voltages by an affine transformation including a $45^\circ$ rotation. The purpose of this rotation is to simplify the slope ranges of the lead-transition segments.

For each $i$0, the parametrization includes the $i$1-intercepts $i$2, the slopes $i$3, and Bézier anchors $i$4 and $i$5 that define the curved portions around the triple points. The length of the purely linear part of each lead-to-dot segment is given as

$i$6

and the number of triple points or Bézier curves in a transition is stated to scale with electron number as

$i$7

Occupation is then reconstructed from the geometry of total-charge and interdot transitions. The region between $i$8 and $i$9 contains exactly $tct_i$0 electrons. Within that region, the division of charge between the two dots is determined using interdot lines $tct_i$1, constructed from the relevant triple points, together with sigmoid functions orthogonal to those lines. The superposition of these sigmoids gives the occupation $tct_i$2 of dot 1, while the second occupation is constrained by total charge as $tct_i$3 (Hader et al., 11 Aug 2025).

This geometric scheme is deliberately non-microscopic. It does not explicitly calculate energies, but it reproduces experimentally relevant layout and curvature, including rounded triple points. This suggests that SimCATS is intended to be calibrated from images and sensor traces rather than from a full electrostatic device model.

4. Ideal CSD generation, sensor modeling, and distortions

Within the framework, “ideal” CSD data are explicitly defined as simulated undisturbed ground-truth data (Hader et al., 11 Aug 2025). In that setting, occupations are computed by the geometric model, sensor response is computed deterministically by the sensor model, and no noise or distortions are applied. The resulting maps contain crisp honeycomb patterns and well-defined boundaries, together with per-pixel labels such as total charge and dot occupations.

The sensor model assumes a nearby sensor dot acting as a charge sensor. It responds both to changes in dot occupation and to applied plunger voltages through cross-capacitance. The model is

$tct_i$4

and

$tct_i$5

Here, $tct_i$6 is the lever arm from dot occupation to sensor potential, $tct_i$7 is the cross-capacitance from plunger gate to sensor potential, and the second equation models the sensor response as a Lorentzian centered at $tct_i$8 with width $tct_i$9. The Lorentzian parameters can be fitted to measured sensor scans, and $i-1$0 and $i-1$1 can be extracted from experimental CSDs by analyzing gradients inside honeycombs and across lead transitions (Hader et al., 11 Aug 2025).

Distortions are grouped into three classes: occupation distortions, sensor-potential distortions, and sensor-response distortions. Occupation distortions include dot jumps and occupation-transition blurring. Sensor-potential distortions include pink noise and random telegraph noise (RTN) acting on $i-1$2. Sensor-response distortions include white noise and additional RTN acting on the final signal $i-1$3 (Hader et al., 11 Aug 2025).

Dot jumps are simulated by shifting blocks of columns or rows in the occupation map, with extension drawn from a geometric distribution and amplitude drawn from a Poisson distribution. Occupation-transition blurring is implemented through a one-dimensional Fermi–Dirac filter kernel along the measurement direction, although for the results reported in the paper this kernel was still approximated by a Gaussian filter. Pink noise is generated following Timmer and Koenig’s method and applied to the sensor potential, which amplifies its visual effect where the Lorentzian response has large slope. RTN is modeled through a geometric distribution for burst extension and a normal distribution for amplitude. White noise is added as Gaussian noise on $i-1$4, with power spectral density written as

$i-1$5

To jointly estimate white and pink noise levels from measurements, the framework fits

$i-1$6

to the high-frequency part of the power spectral density, using constants $i-1$7 and $i-1$8 for Welch’s method in SciPy (Hader et al., 11 Aug 2025).

The overall simulation pipeline is correspondingly staged: ideal occupation is computed, optional broadening and dot jumps are applied, sensor potential is formed, pink noise and RTN may be added at the potential level, the Lorentzian sensor response is evaluated, and final RTN or white noise may be injected at the output stage.

5. Benchmarking automated tuning algorithms

A central stated use case of SimCATS is the development and benchmarking of automated tuning algorithms, including both classical and machine-learning approaches (Hader et al., 11 Aug 2025). Because the framework can vary distortion types and noise strengths systematically, it enables controlled studies of algorithmic sensitivity to measurement quality, sensor parameters, and device irregularities. It also produces ground-truth labels that are typically difficult to obtain at scale from experiments alone.

To validate whether simulated data resemble measured CSDs, the paper uses sample-level metrics drawn from generative-model evaluation: $i-1$9-precision and $i$0-recall. The procedure first embeds CSD images into a feature hypersphere using a convolutional network adapted from MNIST-LeNet and trained with a Deep SVDD objective,

$i$1

where $i$2 is the hypersphere radius, $i$3 its center, $i$4 a balancing factor, $i$5 the number of samples, and $i$6 the learned feature mapping. Precision and recall are then computed through k-nearest-neighbor overlaps between real and simulated samples in that feature space (Hader et al., 11 Aug 2025).

The reported validation results indicate high precision and substantial recall. For a SimCATS dataset with the same size as the experimental test set, precision is reported as approximately $i$7 and recall as approximately $i$8 for $i$9. Expanding the SimCATS dataset to roughly ten times more samples increases recall to approximately $tct_i$0. A comparison with MNIST images yields high precision but tiny recall, around $tct_i$1–$tct_i$2, which is used to illustrate that the metric is not trivial (Hader et al., 11 Aug 2025).

These results are interpreted as evidence that SimCATS occupies a distribution close to measured CSDs while not yet covering all experimental variants. This suggests two complementary uses: first, as a practical source of realistic synthetic data for training and validation; second, as a controlled benchmark for comparing algorithms under identical measurement conditions.

6. Extensibility, limitations, and cross-domain uses of the name

The framework is expressly extensible at the levels of geometry, sensor response, and distortion modeling (Hader et al., 11 Aug 2025). Parameters span transition geometry $tct_i$3, sigmoid widths across interdot transitions, sensor parameters $tct_i$4, and distortion parameters such as jump statistics, noise powers, RTN statistics, and measurement time per voltage point. Because these can be extracted from measured data or set manually, SimCATS can be used both for realistic emulation of a specific platform and for broader robustness sweeps across hypothetical regimes.

The stated limitations are correspondingly clear. The current occupation model is geometric and tailored to double quantum dots; extension to triple or larger arrays would require more complex geometry. The parameters are not directly linked to physical capacitances, tunnel couplings, or generalized Hubbard parameters, which limits the framework’s use for physically predictive simulation. Additional effects, including varying lead-transition lengths across wider voltage ranges, correlations between consecutive CSDs, and correlations across multiple DQDs measured simultaneously, are said not to be fully included yet. Noise estimation routines also have bounded accuracy in certain regimes, such as very low white-noise levels (Hader et al., 11 Aug 2025).

The name “SimCATS” also appears in other research contexts. In the traffic-management paper “Crowdsourcing Autonomous Traffic Simulation,” the formal framework is called CATS, but a structured synthesis explicitly states that one can think of “SimCATS” as the simulation platform implementing the CATS ideas: a microscopic, agent-based traffic simulator in which driver–vehicle agents possess transportation resources, driving credit, behavior types, and Weber–Fechner-based reward or punishment perception, with crowdsourced enforcement by on-board cameras (Wang et al., 2021). In a different domain, the CAT framework for analyzing the consistency–accuracy relation of LLMs is presented as a metric-driven evaluation toolkit, and the structured synthesis is “tailored to someone designing a ‘SimCATS’ system,” where “SimCATS” denotes an envisioned simulation and analysis environment that would generate divergence sets, compute $tct_i$5, $tct_i$6, CAR curves, and the CORE index (Cavalin et al., 26 Nov 2025).

Taken together, these usages indicate that “SimCATS Framework” is not a universally unique label. Its most formal and explicit meaning is the quantum-dot CSD simulator introduced in 2025 (Hader et al., 11 Aug 2025), but the term also functions as an informal or design-oriented shorthand in traffic simulation and LLM evaluation contexts (Wang et al., 2021, Cavalin et al., 26 Nov 2025). This suggests that, in scholarly usage, the term should be disambiguated by domain whenever precision matters.