WorksPhysics and Astronomy
zenDot: An LLM-integrated quantum TCAD platform for semiconductor quantum-device design and optimization automation
Abstract
Semiconductor quantum-device design still lacks an integrated Technology Computer-Aided Design (TCAD)-like environment that connects material geometry, quantum many-body simulation, and automated design. Here we introduce zenDot, a large-language model (LLM)-integrated quantum TCAD platform that links a material-labelled device state to a unified condensed-matter physics toolbox. The device and calculation components are integrated into a desktop workbench, Python API, and an embedded LLM agent, allowing electrostatics, charge and transport characterization, correlated-state calculations, and qubit modelling to be executed within one reproducible environment. We demonstrate zenDot on a Si/SiO2 double quantum dot, where a single device state reproduces the characterization workflow and supports hybrid, tunnel-charge, and singlet-triplet qubit analyses. A platform-level universal-control scan revises the singlet-triplet operating point and reduces the predicted worst-gate infidelity by nearly 30-fold. Beyond analysis, the LLM agent directly operates the same physics environment as human users, proposing design changes, executing registered simulations, and iterating on solver-returned metrics under physics-aware validation. Across three demonstration tasks it completes 18 validated design iterations, including geometry modification followed by a full re-solve from the material stack. zenDot establishes a machine-operable quantum TCAD workflow that connects device physics with LLM-driven design exploration.
Introduction
Technology computer-aided design (TCAD) transformed classical semiconductor development by turning materials, geometry, process conditions, and electrical performance into an iterative computational workflow before fabrication. Semiconductor quantum devices are reaching a similar design bottleneck. Their behaviour depends not only on electrostatics and transport, but also on discrete spectra, tunnel and exchange couplings, many-body states, coherence, initialization, and gate operations. These quantities span different physical descriptions and must be evaluated across a design space that includes material stacks, gate geometry, operating bias, charge configuration, and logical encoding. Conventional device TCAD does not provide this complete route from a physical structure to quantum-device performance. [1, 2, 3, 4, 5, 6, 7, 13, 14, 15, 16, 17]
The required physics is not missing; it is fragmented. Self-consistent quantum-dot electronic structure, charge-stability and transport simulation, correlated-state methods, and non-equilibrium quantum transport have each matured into powerful specialist approaches. In practice, however, they are often connected through separate codes, re-entered parameters, and study-specific scripts. This fragmentation makes it difficult to treat a quantum device as one continuously evolving design object. What is needed is therefore not another isolated solver, but an environment in which a material-defined device can move through electrostatics, characterization, many-body analysis, and qubit modelling without rebuilding the physical state at every stage. [8, 9, 10, 11, 12, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
Large language models add a new reason to solve this integration problem. Tool-using LLMs can already plan multi-step scientific and engineering workflows, while automated methods increasingly assist quantum-dot tuning and charge-state recognition. A quantum-device design agent, however, is useful only when it can act on real simulation tools rather than merely describe what should be calculated. This requires a machine-operable physics environment in which the model can select calculations, change device or control parameters, launch solvers, receive quantitative results, and continue the design loop through the same interfaces available to a human researcher. [32, 33, 34, 35, 36, 37]
Here we present zenDot, an LLM-integrated quantum TCAD platform built around this idea. zenDot combines a material-labelled device representation, a broad condensed-matter physics toolbox, and a common component registry that is exposed identically to the desktop workbench, Python API, and LLM agent. We first establish the platform architecture and its device-to-physics workflow. We then use a double quantum dot to demonstrate characterization from the layer stack and to show that one calibrated device state can support three distinct qubit models. Finally, we place the LLM inside the same computational environment and demonstrate iterative design over both electrical controls and device geometry. The qubit calculations serve as demonstrations of platform breadth; the central result is a quantum-device environment that can be operated by both researchers and language models.
zenDot platform
One device state from geometry to quantum observables
zenDot starts from the physical device. A project script defines materials, layers, gates, reservoirs, and electrical controls, and the resulting device state is shared by every downstream calculation. Material-resolved electrostatics produces the channel confinement, from which localized dots and device-level quantities such as lever arms, capacitance terms, tunnel coupling, and exchange are derived. The same object is then passed to charge, transport, correlated-state, and encoded-qubit calculations. In this way, the platform treats geometry, electrostatics, and quantum observables as successive views of one device rather than as independent inputs assembled after the fact (Figure 1a-Figure 1d).

Figure 1. zenDot connects one device state to human and LLM-operated quantum-device simulation. a, Material-labelled Si/SiO₂ layer stack. b, Channel confinement returned by the electrostatic solution, resolving two minima 42 nm apart. c, Charging matrix derived from the same device state. d, Inter-dot gate control of separation and tunnel coupling. e, Device-to-result workflow shared across the platform. f, Desktop workbench, Python API, and LLM agent access the same component registry. g, Physics-aware LLM execution: proposed design actions are checked, executed by registered solvers, and returned as quantitative results for the next iteration.
The present platform exposes 18 calculation components and 18 visualization components through one registry. The registry spans electrostatics, static and finite-bias charge maps, spectroscopy, correlated states, dynamics, quantum transport, and qubit operations. Each component declares the parameters and device capabilities it requires and the result type it produces. The desktop application builds its calculation panels from this registry, Python calls the same component identifiers, and the LLM agent invokes the same tools. A calculation selected by the agent is therefore the same calculation available to a researcher at the workbench or in a script, not a separate AI-specific approximation (Figure 1e,Figure 1f).
This shared computational path is central to the LLM integration. Numerical outputs carry their units, solver context, and device state so that subsequent calculations, figures, and optimization objectives are connected to the simulation that produced them. The implementation also records the validity range of each component, allowing a broad physics toolbox to coexist without turning unlike observables into interchangeable numbers. These details are transparent to ordinary use but provide the structure required for reproducible multi-stage design.
For automated operation, zenDot adds a physics-aware execution layer on top of the same registry. The LLM can propose a calculation or a bounded change to the declared design variable; zenDot checks that the action is complete and compatible with the current device, executes the registered solver, and returns the measured result to the design loop. This keeps the language model in the role where it is most useful - selecting and sequencing computational actions - while the physics engines remain responsible for the numerical evidence (Figure 1g).
A workbench for human and machine operation
The desktop workbench provides device construction, physical settings, component selection, map configuration, result inspection, export, documentation, and the integrated Assistant within one application (Figure 2). Component availability follows the loaded device and selected backend, while visualization choices follow the result returned by each calculation. Saved projects preserve the device definition and simulation settings needed to reconstruct a study. The Assistant is therefore not an external chatbot attached to a simulator: it is another operating surface of the same quantum-device environment.

Figure 2. The zenDot desktop workbench. a, Device builder and derived confinement. b, Physical and solver settings. c, Registry-generated calculation components and compatible visualizations. d, Charge-stability and transport-map workspace. e, Result visualization and export. f, Integrated documentation. The same registered components are accessible through Python and the LLM Assistant.
Integrated quantum-physics workflow
A condensed-matter toolbox behind one device model
Quantum-device TCAD requires a different physics stack from classical drift-diffusion simulation. zenDot therefore integrates calculations for equilibrium charge and single-electron behaviour, sequential and finite-bias transport, magnetospectroscopy, real-time and quasi-static dynamics, response functions, Wigner-molecule correlations, matrix-product ground states, and non-equilibrium Green-function transport. The transport layer includes surface-Green-function decimation, Fisher-Lee conversion, multiband models, nonorthogonal bases, density-functional NEGF, and time-dependent propagation. These methods are not presented as one universal approximation; they are a toolbox for different regimes of the same device-design problem. [6, 10, 11, 12, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30]
The value of combining these methods is the continuity of the design workflow. A researcher can move from a layer stack to confinement, charge-state structure, tunnelling, spectroscopy, correlated states, and qubit-level observables without rebuilding the device in a second software path. Components become available when their physical requirements are satisfied, and their outputs retain the model and parameter context needed by downstream calculations. Detailed numerical algorithms and approximation ranges are given in the Supporting Information.
From the material stack to device characterization
We demonstrate this workflow on a lateral double quantum dot. The two plunger gates lie on the lower nm plane, while the source, inter-dot, and drain barrier gates lie on the upper nm plane. The 3 nm vertical separation between the gate planes produces, under the chosen controls, two resolved confinement minima separated by 42 nm. Independent calculations of the left and right dots return the same lever arm of 0.28941 eV V, charging energy of 36.73 meV, addition energy of 36.18 meV, and electron factor of 2.000, consistent with the mirror symmetry of the reference device (Figure 3a-Figure 3f,Figure 3k).

Figure 3. Device-level characterization generated from the reference Si/SiO₂ double dot. a-f, Coulomb diamonds, magnetospectroscopy, and Zeeman levels calculated independently for the left and right dots. g, Solved double-well confinement. h, Two-plunger charge-stability map with the selected (1,1) operating point. i, Barrier dependence of exchange and tunnel coupling. j, Empty-cell headroom below the first charging line. k, Left-right consistency of single-dot quantities. l, Even exchange with a cusp-shaped local maximum at .
The same device state then generates the two-dot observables used for subsequent qubit calculations. The charge-stability map resolves four labelled cells including (0,0), with 139 mV of headroom below the first charging line and an interior (1,1) operating point at a common-mode plunger offset of 0.540 V (Figure 3g-Figure 3j). At the reference barrier, the exchange energy is 1.019 µeV and the tunnel coupling is 93.174 µeV. The mirror-symmetric device gives an even exchange function ; in this model is a cusp-shaped local maximum, so the sensitivity at the symmetry point is evaluated from the one-sided slope rather than from a centred difference across the cusp (Figure 3l). Together, these calculations establish the device state inherited by the encoded-qubit demonstrations.
One device state, three qubit models
Cross-encoding demonstration of the platform
The reference double dot provides a compact test of whether one platform can support models that use different electron numbers, operating biases, and Hamiltonians. We therefore evaluate hybrid, tunnel-charge, and singlet-triplet qubits from the same device-defined electrostatic and coupling state. Each model receives the device quantities relevant to its Hamiltonian and returns its own spectrum, dynamics, coherence, thermal, and gate metrics. The purpose is not to rank the three encodings, but to show that a single device model can feed distinct qubit analyses without rebuilding the physical device for each calculation.
Hybrid-qubit demonstration
For the three-electron hybrid qubit, the logical splitting approaches the valley splitting and reaches at the reported operating point, corresponding to . Within the retained charge-noise model and the detuning window, the calculation gives and a quality factor for a rotation. The logical initialization fidelity is 1.000 at 100 mK because the logical splitting is approximately . The complete sweep, thermal populations, and window dependence are shown in Figure 4 and the Supporting Information.

Figure 4. Hybrid-qubit demonstration on the calibrated device. a, Three-electron spectrum. b, Logical splitting approaching the valley splitting. c-e, Noise response, dephasing time, infidelity, and quality factor. f, Dephased logical fringe. g,h, Real-space and depth-resolved density. i,j, Thermal populations and logical initialization. k, First- and second-order amplitude channels. l, Dependence of the reported quality factor on detuning window. At the reported operating point µeV, the frequency is 84.2 GHz, ns, and the logical initialization fidelity is 1.000 at 100 mK.
Tunnel-charge-qubit demonstration
The tunnel-charge qubit uses the device-derived inter-dot tunnel coupling as a two-level anticrossing. At zero detuning, , corresponding to . The same model gives , a modelled infidelity of , and for a rotation. Zero detuning maximizes dephasing time in the retained charge-noise model, while the full gate calculation also captures the finite rotation-axis error. The logical initialization fidelity is 1.000 at 100 mK (Figure 5).

Figure 5. Tunnel-charge-qubit demonstration. a, Charge anticrossing with minimum splitting . b, Charge dipole. c,d, Dephasing and gate-error metrics across detuning. e, Dependence on tunnel coupling. f, Larmor fringe. g,h, Density. i,j, Many-body and logical thermal populations. k, First- and second-order amplitude channels. l, Barrier-controlled gate performance. At zero detuning, the frequency is 45.1 GHz, ns, for a rotation, and the logical initialization fidelity is 1.000 at 100 mK.
Singlet-triplet-qubit demonstration
The singlet-triplet calculation provides a more consequential design example. With a 25 mT dot-to-dot field difference, the rotation axis lies 70.6° from and the worse gate of a universal pair has an infidelity of 0.759, even though the isolated -gate error is . When zenDot scans the Zeeman gradient against the complete universal-control objective, it instead selects (1.90 mT). At this point the splitting is with , the frequency is 252.2 MHz, , and the predicted worst-gate infidelity falls to 0.0259 (Figure 6).

Figure 6. Singlet-triplet-qubit demonstration and platform-level control scan. a, Two-electron spectrum and logical splitting . b,c, Charge-noise suppression and rotation-axis angle. d, Worst-gate infidelity versus Zeeman gradient. e, Dephasing and gate error versus exchange. f, Dephased oscillation. g,h, Density. i,j, Thermal populations. k, Noise channels. l, Initialization versus exchange. The selected operating point gives , 252.2 MHz, ns, and logical initialization fidelity 0.530 at 100 mK.
This approximately 29-fold change in the worst-gate error illustrates the role of the platform rather than establishing a universal optimum for singlet-triplet qubits. The selected point trades charge-noise suppression against coherent axis error, and zenDot exposes both quantities within the same calculation. More broadly, the three encoding branches show that speed, coherence, initialization, and gate error can be evaluated from one device state while remaining tied to the Hamiltonian that defines each metric.
LLM-driven design exploration
The LLM operates the physics platform, not a surrogate
The final demonstration places the LLM inside the design loop. Each study gives the agent a design objective, one allowed variable, its physical bounds, and a simulation budget. The model then proposes the next action using the same registered components available to a human user. zenDot executes the requested calculation, extracts the objective from the solver result, returns the measured outcome, and makes that information available for the next proposal. The agent therefore interacts with a quantum-device simulator as an experimental design environment: it chooses what to change and what to calculate, while the numerical result comes from the underlying physics model (Figure 7a).

Figure 7. LLM-driven quantum-device design in zenDot. a, The agent receives an objective, allowed design variable, bounds, and simulation budget, then operates registered physics components and receives solver-returned metrics. b-d, Hybrid-qubit design trajectory. e-g, Tunnel-charge design trajectory. h-j, Geometry-level singlet-triplet trajectory, including device reconstruction and a full re-solve after each geometric change. k, Sampled locations within the declared design ranges. l, Execution status of the 18 completed iterations. m, Comparison with landscapes scanned using the same physics components. Full prompts, replies, numerical histories, and validation details are provided in the Supporting Information.
Across three demonstration tasks, the LLM completes 18 solver-validated design iterations. In the hybrid-qubit task, six barrier proposals increase from 161412.67 to 220112.30, with the best sampled point at -0.300 V. In the tunnel-charge task, six proposals increase from 225.07 to 809.67, with the best point at -0.028906 V. These campaigns are intentionally small, but they demonstrate a complete closed loop from natural-language planning to parameter change, physics execution, quantitative feedback, and another model action (Figure 7b-Figure 7g).
The third task moves beyond tuning an electrical control. The LLM varies the inter-dot gate length, causing zenDot to modify the physical geometry and rebuild the device before evaluating the singlet-triplet objective. Six geometry-level iterations reduce the measured infidelity from to at a gate length of 22.5 nm (Figure 7h-Figure 7j). Although the numerical improvement in this particular landscape is small, the workflow is the important result: the language model is not optimizing a fitted response surface, but editing a device and triggering a fresh physics calculation from the material stack. Scanned landscapes and complete agent histories are retained in the Supporting Information for quantitative comparison of the sampled trajectories.
Discussion
zenDot reframes semiconductor quantum-device simulation as a connected design environment rather than a collection of independent calculations. The platform starts from a material-defined device and carries that state through electrostatics, charge and transport characterization, many-body physics, and encoded-qubit analysis. The same computational components are accessible through a graphical workbench, Python, and an LLM interface. This combination is important because machine operation is not added after the physics workflow is built; it is a native mode of operating the same workflow.
The multi-encoding demonstration shows why this integration matters. Hybrid, tunnel-charge, and singlet-triplet qubits use different Hamiltonians and physical scales, yet all can be evaluated from the same reference device. The singlet-triplet example is especially illustrative: a control choice that appears favourable when viewed through charge-noise suppression alone changes when universal control is evaluated, reducing the predicted worst-gate infidelity from 0.759 to 0.0259. The specific numbers are conditional on the modelled device, but the broader capability is transferable - coupled device and qubit metrics can be explored without leaving the common design state.
The LLM integration extends this idea from interactive simulation to machine-operated design. In zenDot, the model can select calculations, modify electrical controls, edit geometry, receive quantitative solver feedback, and continue the search. The geometry experiment is particularly important because it crosses the boundary between tuning and design: changing the gate length forces the platform to reconstruct and re-solve the device rather than querying a precomputed surrogate. This is the route by which language models can become useful participants in quantum-device engineering - not by replacing physical modelling, but by navigating it.
The present agent runs are demonstrations of this operating model rather than benchmarks of optimization intelligence. They use one local language model and six evaluations per task, and the sampled trajectories are not expected to establish optimizer superiority. Likewise, the reference device is an internally consistent model rather than a fit to a fabricated device, relaxation is not included in the quoted gate fidelities, and material and geometry uncertainties are not yet propagated through the full workflow. These limitations define the next validation steps without changing the platform-level result demonstrated here.
Experimental calibration, uncertainty propagation, cross-backend comparison, and larger matched-budget agent studies can all be added without changing the architecture. A calibrated device can replace the present reference state; alternative physics backends can be evaluated through the same component interface; and different agent or deterministic strategies can operate under identical design variables and objectives. zenDot therefore provides not only a toolkit for the calculations reported here, but a common substrate for increasingly automated and experimentally grounded quantum-device design.
Conclusion
zenDot integrates semiconductor quantum-device construction, condensed-matter simulation, qubit analysis, and LLM-driven design within one quantum TCAD platform. A single material-labelled device state can be carried from geometry to experimental observables and distinct qubit models, while the same physics components can be operated through the desktop workbench, Python, or an embedded language-model agent. The demonstrations show the breadth of the workflow, and the 18 agent iterations show that an LLM can act directly on the physics environment, including modifying device geometry and triggering a full re-solve. By making quantum-device simulation machine-operable without separating automation from the underlying physics, zenDot provides a foundation for autonomous, experimentally calibrated quantum-device engineering.
Methods
Platform architecture and component registry
A zenDot project contains a device source and a typed settings object. The device source defines material layers, gates, reservoirs, and electrical control names; the settings specify charge sectors, spin representation, solver backend, temperature, fields, noise parameters, and scan controls. Loading a project constructs the common device state consumed by registered calculation components. Each component declares a stable identifier, parameter schema, device and backend requirements, returned result type, and supported visualizations. Completed results retain component identity, parameters, units, solver metadata, and device provenance. The full component inventory and result-field definitions are provided in the Supporting Information.
Device and electrostatics
The reference device is a lateral double dot with an 8 nm Si channel, oxide, and aluminium-oxide layers. Plunger gates lie at nm; source, inter-dot, and drain barrier gates lie at nm. The electrostatic calculation applies material permittivities and gate boundary conditions to the scripted geometry. Localized dots are identified from the resolved channel confinement, after which lever arms, charging terms, tunnel coupling, and exchange are computed from the same device state.
eq:1. Material-resolved electrostatics
Charge and transport
Charge-stability maps use a constant-interaction free energy evaluated over integer charge states. Sequential current is calculated with a first-order transition-rate master equation. The corresponding transport results are interpreted within this sequential-tunnelling approximation. [3, 7, 30]
(1). Constant-interaction free energy
(2). Sequential master equation and lead current
Encoded-qubit models
The hybrid-qubit component diagonalizes the retained three-electron logical subspace using the device-derived tunnel parameters and a declared valley splitting. The tunnel-charge component uses a two-level charge Hamiltonian with the device-derived tunnel coupling. The singlet-triplet component combines device-derived exchange with a Zeeman-energy difference . Each model computes its own spectrum, frequency, coherence, thermal populations, dynamics, and gate metrics from the corresponding returned arrays.
(3). Tunnel-charge Hamiltonian
(4). Singlet-triplet Hamiltonian
(5). Frequency, gate time, and quality factor
Noise, thermal state, and declared model parameters
Four parameters enter the encoded stage in addition to quantities obtained directly from the electrostatic device state: a charge-noise amplitude of 1.5 at 1 Hz, a residual-gradient dephasing time of 1 µs, an electron temperature of 100 mK, and the valley splitting. zenDot evaluates the valley splitting using the 8 nm channel width and an assumed vertical channel field of 10 mV nm rather than the field returned by the Poisson solution. First- and second-order splitting sensitivities are evaluated numerically at the selected operating point. Thermal populations use the calculated energy levels and retained degeneracies. The quality factor is , where depends on the rotation angle ; the quoted hybrid value uses a rotation, whereas the tunnel-charge and singlet-triplet values use rotations.
(6). First- and second-order splitting noise
(7). Thermal occupation
LLM-driven design execution
Each LLM design task specifies an objective, metric path, optimization direction, one allowed design variable, physical bounds, and a completed-simulation budget. The model returns a structured action that selects a registered component and proposes the next parameter value. zenDot verifies that the action is syntactically complete and compatible with the declared task and current device before execution. The selected component then performs the physics calculation, and the objective is read from the returned result rather than from the model text. The measured history is passed back to the design loop for subsequent proposals. Complete task declarations, prompts, replies, execution records, and pseudocode are provided in the Supporting Information.
(8). Bounded optimization over valid device states
Data availability
The source data underlying all seven figures, including device, characterization, encoding, state, and optimization records, are provided with this paper as a unified Source Data package. Saved dialogue records and cross-study validation summaries are included in the Supporting Information.
Code availability
The zenDot source code, device definitions, calculation stages, and figure-building scripts used in this study are being prepared for commercial development and are not publicly available. Non-commercial academic evaluation access may be requested from the corresponding author and may be provided under applicable confidentiality, evaluation, or licensing terms. Confidential access can be made available to editors and reviewers during peer review where required.
Acknowledgements
The author acknowledges the use of OpenAI Codex and Anthropic Claude Code for software-development assistance, code review, documentation, and language editing. All AI-assisted outputs were reviewed, tested, and validated by the author, who takes responsibility for the final code and manuscript.
References
- [1]D. Loss and D. P. DiVincenzo. Quantum computation with quantum dots. Physical Review A 57(1), 120–126 (1998). https://doi.org/10.1103/PhysRevA.57.120
- [2]R. Hanson, L. P. Kouwenhoven, J. R. Petta, S. Tarucha and L. M. K. Vandersypen. Spins in few-electron quantum dots. Reviews of Modern Physics 79(4), 1217–1265 (2007). https://doi.org/10.1103/RevModPhys.79.1217
- [3]W. G. van der Wiel, S. De Franceschi, J. M. Elzerman, T. Fujisawa, S. Tarucha and L. P. Kouwenhoven. Electron transport through double quantum dots. Reviews of Modern Physics 75(1), 1–22 (2002). https://doi.org/10.1103/RevModPhys.75.1
- [4]J. M. Elzerman, R. Hanson, L. H. Willems van Bevereen, B. Witkamp, L. M. K. Vandersypen and L. P. Kouwenhoven. Single-shot read-out of an individual electron spin in a quantum dot. Nature 430(6998), 431–435 (2004). https://doi.org/10.1038/nature02693
- [5]J. R. Petta, A. C. Johnson, J. M. Taylor, E. A. Laird, A. Yacoby, M. D. Lukin, C. M. Marcus, M. P. Hanson and A. C. Gossard. Coherent manipulation of coupled electron spins in semiconductor quantum dots. Science 309(5744), 2180–2184 (2005). https://doi.org/10.1126/science.1116955
- [6]K. K. Likharev. Single-electron devices and their applications. Proceedings of the IEEE 87(4), 606–632 (1999). https://doi.org/10.1109/5.752518
- [7]C. W. J. Beenakker. Theory of Coulomb-blockade oscillations in the conductance of a quantum dot. Physical Review B 44(4), 1646–1656 (1991). https://doi.org/10.1103/PhysRevB.44.1646
- [8]M. Stopa. Quantum dot self-consistent electronic structure and the Coulomb blockade. Physical Review B 54(19), 13767–13783 (1996). https://doi.org/10.1103/PhysRevB.54.13767
- [9]S. M. Reimann and M. Manninen. Electronic structure of quantum dots. Reviews of Modern Physics 74(4), 1283–1342 (2002). https://doi.org/10.1103/RevModPhys.74.1283
- [10]R. Egger, W. Häusler, C. H. Mak and H. Grabert. Crossover from Fermi liquid to Wigner molecule behavior in quantum dots. Physical Review Letters 82(16), 3320–3323 (1999). https://doi.org/10.1103/PhysRevLett.82.3320
- [11]I. Shapir, A. Hamo, S. Pecker, C. P. Moca, Ö. Legeza, G. Zaránd and S. Ilani. Imaging the electronic Wigner crystal in one dimension. Science 364(6443), 870–875 (2019). https://doi.org/10.1126/science.aat0905
- [12]W. Jang, J. Kim, J. Park, G. Kim, M.-K. Cho, H. Jang, S. Sim, B. Kang, H. Jung, V. Umansky and D. Kim. Wigner-molecularization-enabled dynamic nuclear polarization. Nature Communications 14(1), 2948 (2023). https://doi.org/10.1038/s41467-023-38649-5
- [13]J. Yoneda, K. Takeda, T. Otsuka, T. Nakajima, M. R. Delbecq, G. Allison, T. Honda, T. Kodera, S. Oda, Y. Hoshi, N. Usami, K. M. Itoh and S. Tarucha. A quantum-dot spin qubit with coherence limited by charge noise and fidelity higher than 99.9%. Nature Nanotechnology 13(2), 102–106 (2018). https://doi.org/10.1038/s41565-017-0014-x
- [14]M. Veldhorst, C. H. Yang, J. C. C. Hwang, W. Huang, J. P. Dehollain, J. T. Muhonen, S. Simmons, A. Laucht, F. E. Hudson, K. M. Itoh, A. Morello and A. S. Dzurak. A two-qubit logic gate in silicon. Nature 526(7573), 410–414 (2015). https://doi.org/10.1038/nature15263
- [15]X. Xue, M. Russ, N. Samkharadze, B. Undseth, A. Sammak, G. Scappucci, and L. M. K. Vandersypen. Quantum logic with spin qubits crossing the surface code threshold. Nature 601(7893), 343–347 (2022). https://doi.org/10.1038/s41586-021-04273-w
- [16]W. Huang, C. H. Yang, K. W. Chan, T. Tanttu, B. Hensen, R. C. C. Leon, M. A. Fogarty, J. C. C. Hwang, F. E. Hudson, K. M. Itoh, A. Morello, A. Laucht, and A. S. Dzurak. Fidelity benchmarks for two-qubit gates in silicon. Nature 569(7757), 532–536 (2019). https://doi.org/10.1038/s41586-019-1197-0
- [17]A. J. Weinstein, M. D. Reed, A. M. Jones, R. W. Andrews, D. Barnes, J. Z. Blumoff, L. E. Euliss, K. Eng, B. H. Fong, S. D. Ha, D. R. Hulbert, C. A. C. Jackson, M. Jura, T. E. Keating, J. Kerckhoff, A. A. Kiselev, J. Matten, G. Sabbir, A. Smith, J. Wright, M. T. Rakher, T. D. Ladd, and M. G. Borselli. Universal logic with encoded spin qubits in silicon. Nature 615(7954), 817–822 (2023). https://doi.org/10.1038/s41586-023-05777-3
- [18]Y. Meir and N. S. Wingreen. Landauer formula for the current through an interacting electron region. Physical Review Letters 68(16), 2512–2515 (1992). https://doi.org/10.1103/PhysRevLett.68.2512
- [19]A.-P. Jauho, N. S. Wingreen, and Y. Meir. Time-dependent transport in interacting and noninteracting resonant-tunneling systems. Physical Review B 50(8), 5528–5544 (1994). https://doi.org/10.1103/PhysRevB.50.5528
- [20]R. Lake, G. Klimeck, R. C. Bowen, and D. Jovanovic. Single and multiband modeling of quantum electron transport through layered semiconductor devices. Journal of Applied Physics 81(12), 7845–7869 (1997). https://doi.org/10.1063/1.365394
- [21]M. Brandbyge, J.-L. Mozos, P. Ordejón, J. Taylor, and K. Stokbro. Density-functional method for nonequilibrium electron transport. Physical Review B 65(16), 165401 (2002). https://doi.org/10.1103/PhysRevB.65.165401
- [22]K. S. Thygesen. Electron transport through an interacting region: The case of a nonorthogonal basis set. Physical Review B 73(3), 035309 (2006). https://doi.org/10.1103/PhysRevB.73.035309
- [23]M. P. López Sancho, J. M. López Sancho, and J. Rubio. Quick iterative scheme for the calculation of transfer matrices: application to Mo(100). Journal of Physics F: Metal Physics 14(5), 1205–1215 (1984). https://doi.org/10.1088/0305-4608/14/5/016
- [24]M. P. López Sancho, J. M. López Sancho, J. M. L. Sancho, and J. Rubio. Highly convergent schemes for the calculation of bulk and surface Green functions. Journal of Physics F: Metal Physics 15(4), 851–858 (1985). https://doi.org/10.1088/0305-4608/15/4/009
- [25]D. S. Fisher and P. A. Lee. Relation between conductivity and transmission matrix. Physical Review B 23(12), 6851–6854 (1981). https://doi.org/10.1103/PhysRevB.23.6851
- [26]A. Croy and U. Saalmann. Propagation scheme for nonequilibrium dynamics of electron transport in nanoscale devices. Physical Review B 80(24), 245311 (2009). https://doi.org/10.1103/PhysRevB.80.245311
- [27]N. S. Wingreen, A.-P. Jauho, and Y. Meir. Time-dependent transport through a mesoscopic structure. Physical Review B 48(11), 8487–8490 (1993). https://doi.org/10.1103/PhysRevB.48.8487
- [28]C. W. Groth, M. Wimmer, A. R. Akhmerov and X. Waintal. Kwant: a software package for quantum transport. New Journal of Physics 16(6), 063065 (2014). https://doi.org/10.1088/1367-2630/16/6/063065 zenDot | LLM-integrated quantum TCAD platform Preliminary Version
- [29]T. Kloss, J. Weston, B. Gaury, B. Rossignol, C. Groth and X. Waintal. Tkwant: a software package for time-dependent quantum transport. New Journal of Physics 23(2), 023025 (2021). https://doi.org/10.1088/1367-2630/abddf7
- [30]G. Kiršanskas, J. N. Pedersen, O. Karlström, M. Leijnse and A. Wacker. QmeQ 1.0: An open-source Python package for calculations of transport through quantum dot devices. Computer Physics Communications 221, 317–342 (2017). https://doi.org/10.1016/j.cpc.2017.07.024
- [31]D. L. Buterakos, S. S. Kalantre, J. Ziegler, J. M. Taylor and J. P. Zwolak. QDFlow: A Python package for physics simulations of quantum dot devices. SciPost Physics Codebases 65 (2026). https://doi.org/10.21468/scipostphyscodeb.65
- [32]J. P. Zwolak and J. M. Taylor. Colloquium: Advances in automation of quantum dot devices control. Reviews of Modern Physics 95(1), 011006 (2023). https://doi.org/10.1103/revmodphys.95.011006
- [33]J. Darulová, S. J. Pauka, N. Wiebe, K. W. Chan, G. C. Gardener, M. J. Manfra, M. C. Cassidy and M. Troyer. Autonomous tuning and charge-state detection of gate-defined quantum dots. Physical Review Applied 13(5), 054005 (2020). https://doi.org/10.1103/physrevapplied.13.054005
- [34]F. Hader, S. Fleitmann, J. Vogelbruch, L. Geck and S. van Waasen. Simulation of charge stability diagrams for automated tuning solutions (SimCATS). IEEE Transactions on Quantum Engineering 5, 1–14 (2024). https://doi.org/10.1109/tqe.2024.3445967
- [35]H. Wu, Z. He, X. Zhang, X. Yao, S. Zheng, H. Zheng and B. Yu. ChatEDA: A large language model powered autonomous agent for EDA. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 43(10), 3184–3197 (2024). https://doi.org/10.1109/tcad.2024.3383347
- [36]D. A. Boiko, R. MacKnight, B. Kline and G. Gomes. Autonomous chemical research with large language models. Nature 624(7992), 570–578 (2023). https://doi.org/10.1038/s41586-023-06792-0
- [37]A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White and P. Schwaller. Augmenting large language models with chemistry tools. Nature Machine Intelligence 6(5), 525–535 (2024). https://doi.org/10.1038/s42256-024-00832-8