Executive Summary: FNO
1. Governing Physical and Computational Mechanism: This monograph establishes a unified, differentiable inverse co-design framework for electrostatically defined semiconductor quantum-dot devices, authored by Marek Rychetsky, Valentin Duruisseaux, Freya Shah, David L. Craig, Ming Ni, and Georgios Katsaros (California Institute of Technology and University of California, Berkeley; arXiv:2611.06994). The framework solves the fundamental inverse problem of quantum device design by establishing a direct causal link between lithographic gate electrode geometries, applied control voltages, and the resulting electronic confinement profiles. High-fidelity training data generated via a semiclassical Thomas–Fermi electrostatic simulator conditions a Fourier Neural Operator (FNO) surrogate model. The FNO operates as a resolution-independent operator, learning continuous mappings from top-gate layout topologies and boundary potential fields to two-dimensional electron charge-density fields ($\rho(x, y)$).
2. Quantitative Benchmarks and Performance Metrics: Empirical validation confirms that the FNO surrogate reproduces complex two-dimensional charge-density distributions with a relative $L^2$ test error of just $0.6\%$. Executed on GPU architecture, the neural operator achieves a 17,000-fold acceleration over the standard CPU-based Thomas–Fermi numerical simulator, reducing evaluation times to the millisecond scale while retaining exact computational differentiability. When embedded in an end-to-end gradient-based optimization framework under strict lithographic fabrication constraints, the FNO successfully synthesizes customized confinement landscapes—generating both vertically and horizontally elongated double-quantum-dot geometries from identical initial conditions. Batched inference scales parameter exploration across dense optimization spaces.
3. Strategic Significance and Technological Implications: By shifting quantum-dot architecture development from empirical fabrication iteration to predictive inverse co-design, this approach bypasses traditional heuristic design bottlenecks. Systematically engineering spatial charge distributions prior to physical lithography offers precise deterministic control over exchange coupling, tunnel barriers, and qubit operational properties, offering a scalable methodology for spin-based quantum computing architectures.
Theoretical Framework & Mechanisms: FNO
Semiclassical Electrostatics and Charge-Density Formulation
The electrostatic state of electrostatically defined semiconductor quantum-dot devices is governed by the self-consistent interaction between applied gate voltages, spatial electrode geometry, and charge carrier dynamics within a two-dimensional electron gas (2DEG). The governing partial differential equation for the electrostatic potential field $V(\mathbf{r})$ in the device domain $\Omega \subset \mathbb{R}^2$ is expressed via the non-homogeneous Poisson equation:
$$\nabla \cdot \left( \epsilon(\mathbf{r}) \nabla V(\mathbf{r}) \right) = -e \left( N_D^+(\mathbf{r}) - n(\mathbf{r}; V) \right)$$where $\epsilon(\mathbf{r})$ represents the spatial dielectric permittivity profile of the semiconductor heterostructure, $N_D^+(\mathbf{r})$ denotes the ionized donor concentration, $e$ is the elementary charge, and $n(\mathbf{r}; V)$ is the 2D electron charge density. Under the semiclassical Thomas–Fermi approximation at sub-Kelvin operational temperatures, the carrier concentration $n(\mathbf{r})$ is explicitly coupled to the local electrostatic potential $V(\mathbf{r})$ and Fermi level $E_F$ according to:
$$n(\mathbf{r}) = \frac{m^*}{\pi \hbar^2} \max\left( 0, E_F - e V(\mathbf{r}) \right)$$where $m^*$ is the effective carrier mass and $\hbar$ is the reduced Planck constant. The non-linear zero-floor operator $\max(0, \cdot)$ introduces sharp boundary discontinuities at the depleted quantum-dot interfaces, making conventional CPU-based iterative finite-element Poisson–Thomas–Fermi solvers computationally intensive during high-dimensional parameter sweeps.
Fourier Neural Operator Architecture on Infinite-Dimensional Function Spaces
To overcome the computational barrier of full numerical electrostatics, the continuous forward mapping from control gate configurations to the spatial confinement field is formulated as a parametric operator mapping between infinite-dimensional Banach spaces $\mathcal{G}_{\theta}: \mathcal{A} \to \mathcal{U}$. Here, the input function $a(\mathbf{r}) \in \mathcal{A}$ parametrizes both the lithographic gate geometry $\boldsymbol{\theta}_g$ and the applied voltage vector $\mathbf{V}_g$, while $u(\mathbf{r}) = n(\mathbf{r}) \in \mathcal{U}$ represents the resulting 2D charge density target field.
The Fourier Neural Operator (FNO) architecture processes spatial features through $L$ sequential neural operator updates. For layer index $t \in \{0, \dots, L-1\}$, the feature map update $v_t(\mathbf{x}) \in \mathbb{R}^{d_v}$ is defined by:
$$v_{t+1}(\mathbf{x}) = \sigma \left( W v_t(\mathbf{x}) + \left( \mathcal{K}(\phi) v_t \right)(\mathbf{x}) \right)$$where $W \in \mathbb{R}^{d_v \times d_v}$ is a local linear transformation, $\sigma$ is a non-linear activation function, and $\mathcal{K}(\phi)$ is a non-local kernel integration operator defined in Fourier space:
$$\left( \mathcal{K}(\phi) v_t \right)(\mathbf{x}) = \mathcal{F}^{-1} \left( R_\phi \cdot \left(\mathcal{F} v_t\right) \right)(\mathbf{x})$$Here, $\mathcal{F}$ and $\mathcal{F}^{-1}$ denote the forward and inverse Fast Fourier Transforms, respectively, and $R_\phi$ represents a parameterized complex-valued tensor acting on truncated lower Fourier modes up to a maximum wavenumber $k_{\text{max}}$. Because spatial convolution reduces to element-wise complex multiplication in the spectral domain, the FNO achieves grid-independent resolution invariance, enabling inference speed-ups of approximately 17,000-fold over CPU baseline physics solvers while retaining a relative $L^2$ prediction error bounded at $0.6\%$.
Differentiable Joint Inverse Co-Design Mechanics
The joint optimization of gate electrode layout $\boldsymbol{\theta}_g$ and control voltages $\mathbf{V}_g$ is executed by leveraging the end-to-end differentiability of the operator network $\mathcal{G}_{\theta}$. The inverse co-design task is framed as a continuous optimization problem governing target spatial confinement $n_{\text{target}}(\mathbf{r})$:
$$\min_{\boldsymbol{\theta}_g, \mathbf{V}_g} \mathcal{L}_{\text{total}} = \frac{\| \mathcal{G}_{\theta}(\boldsymbol{\theta}_g, \mathbf{V}_g) - n_{\text{target}}(\mathbf{r}) \|_{L^2}^2}{\| n_{\text{target}}(\mathbf{r}) \|_{L^2}^2} + \lambda_{\text{fab}} \mathcal{R}_{\text{fab}}(\boldsymbol{\theta}_g)$$where $\mathcal{R}_{\text{fab}}(\boldsymbol{\theta}_g)$ is a differentiable fabrication regularizer imposing physical constraints such as minimum lithographic feature dimensions $d_{\text{min}}$ and boundary curvature bounds $\kappa(\mathbf{s})$ along the electrode perimeter $\partial \Omega$:
$$\mathcal{R}_{\text{fab}}(\boldsymbol{\theta}_g) = \int_{\partial \Omega} \left[ \max\left(0, d_{\text{min}} - \|\nabla \boldsymbol{\theta}_g\|\right)^2 + \gamma |\kappa(\mathbf{s})|^2 \right] d\mathbf{s}$$By computing analytical gradients $\nabla_{\boldsymbol{\theta}_g, \mathbf{V}_g} \mathcal{L}_{\text{total}}$ via backpropagation through the Fourier layers, the gradient-based design loop dynamically reshapes gate boundaries and voltage bias distributions, enabling deterministic synthesis of specialized target geometries—including horizontally and vertically elongated double-dot confinement states—from unified initial conditions.
Empirical Evidence & Methodology: FNO
Surrogate Accuracy and Computational Acceleration Benchmarks
The empirical validation of the Fourier Neural Operator (FNO) surrogate framework rests upon its ability to approximate the high-dimensional nonlinear charge-density mapping derived from the semiclassical Thomas–Fermi electrostatic simulator while drastically reducing evaluation latency. Quantitative assessment of the predictive fidelity reveals that the trained FNO achieves a relative $L^2$ test error of precisely $0.6\%$, demonstrating exceptional accuracy across heterogeneous gate layout topographies and potential landscapes. This sub-percent error bound ensures that fine spatial features of the two-dimensional electron gas (2DEG) and subtle quantum confinement boundaries are faithfully preserved during iterative optimization.
From a computational efficiency perspective, the operator learning approach fundamentally transforms the throughput of inverse quantum-dot synthesis. Traditional CPU-based self-consistent solver routines present a severe bottleneck in gradient-based design loops. In contrast, the GPU-accelerated FNO surrogate executes millisecond-scale forward and backward differentiable evaluations. Benchmark comparisons establish a $17,000$-fold acceleration for single-device layout evaluations relative to standard CPU-bound physics simulators. Furthermore, the structural capability for parallelized tensor operations allows batched inference routines to yield additional compound throughput gains, enabling exhaustive design space exploration that was previously computationally intractable.
Morphological Synthesis and Inverse Co-Design Validation
The practical utility of the differentiable FNO surrogate was evaluated by embedding the model within a gradient-based optimization pipeline constrained by realistic nanofabrication lithography boundaries. Starting from identical initial electrostatic conditions, the inverse co-design architecture was tasked with targeting distinct spatial confinement profiles. The empirical results demonstrate robust convergence to structurally diverse double-quantum-dot architectures:
- Vertically Elongated Double-Dot Geometries: The co-design loop successfully synthesized non-trivial gate topographies and potential biases that deform the spatial electron density along the transverse axis, modifying inter-dot coupling parameters.
- Horizontally Elongated Double-Dot Geometries: Under modified target field specifications, the system converged to longitudinal confinement profiles, showcasing precise control over both quantum-dot shape and spatial centroid position prior to physical device fabrication.
Formal Research Attribution
The empirical methodologies, numerical datasets, computational benchmarks, and architectural paradigms discussed in this chapter are directly derived from the foundational literature documented below.
Primary Citation & Institutional Metadata:
Title: FNO-Based Inverse Co-Design of Semiconducting Quantum-Dot Devices
Authors: Marek Rychetsky, Valentin Duruisseaux, Freya Shah, David L. Craig, Ming Ni, Georgios Katsaros
Institutional Affiliations: California Institute of Technology; University of California, Berkeley
Publication Venue & Identifier: arXiv Preprint (Quantum Physics, quant-ph), arXiv:2611.06994
Canonical URI / DOI: https://arxiv.org/abs/2611.06994 (📄 DOI: 10.48550/arXiv.2611.06994)
Scientific Impact & Future Horizons: FNO
While the Fourier Neural Operator (FNO) surrogate framework exhibits exceptional millisecond-scale evaluation speeds and continuous differentiability for two-dimensional electrostatic potential fields, critical methodological frontiers remain to be traversed. Currently, the inverse co-design framework relies on a semiclassical Thomas–Fermi approximation mapped across planar gate topographies. Extending this paradigm to capture full three-dimensional quantum confinement requires integrating self-consistent Poisson–Schrödinger dynamics directly into the operator learning pipeline. Furthermore, physical device realizations inevitably introduce disorder potentials stemming from interfacial roughness, stray charge traps, and random dopant fluctuations. Incorporating stochastic perturbation models and robust optimization techniques into the differential inverse design loop will ensure that synthesized gate layouts maintain target quantum-dot geometries under non-ideal experimental conditions.
Beyond single- and double-quantum-dot architectures, the translational value of FNO-driven inverse co-design lies in its capacity to scale towards large-scale spin-qubit processor tiles. In high-density two-dimensional crossbar arrays, capacitive crosstalk and tight geometric pinout constraints severely hamper manual gate design. By formulating multi-objective loss functions that encode gate voltage thermal dissipation budgets, layout lithographic fabrication limits, and precise inter-dot tunnel coupling metrics, the FNO framework enables automated synthesis of complex quantum dot arrays. This paradigm shifts quantum hardware engineering from heuristic trial-and-error prototyping to automated, physics-informed inverse synthesis, significantly shortening iteration cycles for fault-tolerant spin-qubit devices.
Core Takeaways
- Fundamental Mechanism: Infinite-dimensional operator learning via Fourier Neural Operators maps continuous electrostatic boundary conditions to spatial charge density distributions, enabling exact gradient propagation for simultaneous gate geometry and bias voltage co-optimization.
- Real-World Value: Eliminates empirical fabrication-and-test design loops by automating the synthesis of custom quantum-dot confinement landscapes, accelerating the development of scalable, defect-tolerant silicon and germanium spin-qubit architectures.
Applications & Future Outlook
Immediate application paths include the automated optimization of high-density 2D semiconductor spin-qubit arrays, automated tuning of tunnel couplings, and integrated crosstalk suppression. Key open challenges involve expanding operator training to full 3D self-consistent Poisson–Schrödinger systems and accounting for stochastic substrate disorder.
References
- M. Rychetsky, V. Duruisseaux, F. Shah, D. L. Craig, M. Ni, and G. Katsaros, "FNO-Based Inverse Co-Design of Semiconducting Quantum-Dot Devices," arXiv preprint arXiv:2611.06994, 2026. 📄 DOI: 10.48550/arXiv.2611.06994.
- Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, "Fourier Neural Operator for Parametric Partial Differential Equations," in International Conference on Learning Representations (ICLR), 2021.
- L. M. K. Vandersypen, H. Bluhm, J. S. Clarke, A. S. Dzurak, R. Ishihara, A. Morello, D. J. Reilly, L. R. Schreiber, and M. Veldhorst, "Interfacing spin qubits in quantum dots—a quantum computing architecture challenge," npj Quantum Information, vol. 3, no. 1, p. 34, 2017.
💬 Comments