🔬 mesoscale physics

Cross-Platform Autonomous Control of Minimal Kitaev Chains

This paper demonstrates the successful cross-platform autonomous tuning of Poor Man's Majorana zero modes in minimal Kitaev chains by employing a transfer learning approach that trains a convolutional neural network on theoretical models and refines it on experimental data to rapidly converge on optimal electrochemical potentials.

Original authors: David van Driel, Rouven Koch, Vincent P. M. Sietses, Sebastiaan L. D. ten Haaf, Chun-Xiao Liu, Francesco Zatelli, Bart Roovers, Alberto Bordin, Nick van Loo, Guanzhong Wang, Jan Cornelis Wolff, Grzego

Published 2026-06-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: David van Driel, Rouven Koch, Vincent P. M. Sietses, Sebastiaan L. D. ten Haaf, Chun-Xiao Liu, Francesco Zatelli, Bart Roovers, Alberto Bordin, Nick van Loo, Guanzhong Wang, Jan Cornelis Wolff, Grzegorz P. Mazur, Tom Dvir, Ivan Kulesh, Qingzhen Wang, A. Mert Bozkurt, Sasa Gazibegovic, Ghada Badawy, Erik P. A. M. Bakkers, Michael Wimmer, Srijit Goswami, Jose L. Lado, Leo P. Kouwenhoven, Eliska Greplova

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to tune a very complex musical instrument, like a giant, futuristic harp made of electricity. This instrument, called a Kitaev chain, is designed to play a very special, rare note known as a "Poor Man's Majorana" mode. This note is special because it could one day help build super-powerful quantum computers.

However, tuning this instrument is incredibly difficult. It has many knobs (called gates) that control the flow of electricity, and if you turn them even slightly wrong, the special note disappears. Traditionally, a human expert would have to sit there for hours, turning knobs, listening to the "music" (measuring electrical signals), and guessing which way to turn next. It's slow, frustrating, and requires a lot of trial and error.

The "Robot Tuner" Solution
The researchers in this paper built a "robot tuner" using Artificial Intelligence (AI) to do this job automatically. Here is how they did it, using a clever trick called Cross-Platform Transfer Learning:

Learning the Theory (The Simulator): First, they taught the AI how the instrument should work using a perfect computer simulation. Think of this like a music student studying sheet music and theory in a quiet classroom. The AI learned what the "perfect note" looks like in a perfect world.
The "Practice" Instrument (Device B): Next, they tested the AI on a real, but slightly different, version of the instrument (a flat chip made of a special material). They didn't give the AI all the data; they just showed it a few examples of what the real instrument sounded like. This was like the student playing a few songs on a real piano to get used to the feel of the keys.
The "Grand Performance" (Device A): Finally, they sent the AI to tune the actual target instrument (a tiny wire device). Crucially, the AI had never seen this specific wire before. It had to use what it learned from the theory and the practice chip to figure out how to tune the new wire.

How the AI "Hears" the Note
The AI doesn't listen with ears; it looks at a "map" of the electricity called a Charge Stability Diagram.

Imagine this map as a landscape with hills and valleys.
The AI uses a special eye (a Convolutional Neural Network) to scan this map. It looks for a specific pattern: a "cross" shape where two different electrical processes balance each other perfectly.
When it finds this cross, it knows it has hit the "Sweet Spot"—the perfect place where the special quantum note exists.

The Results
The AI didn't just guess; it used a mathematical method called "gradient descent" (think of it as a hiker feeling their way down a hill to find the lowest valley) to adjust the knobs automatically.

Speed: In about 45 minutes, the AI successfully found the sweet spot. A human might take much longer.
Accuracy: In about 68% of attempts, the AI found the spot within a tiny margin of error (±1.5 mV). In 81% of attempts, it was within a slightly larger margin (±4.5 mV).
Versatility: It worked even when they turned on a magnetic field, which makes the "instrument" much harder to tune.

The Big Picture
The paper concludes that this method is a major step forward. It proves that an AI trained on a computer and a small practice chip can successfully tune a completely different, complex device without needing to be retrained on that specific device first.

The researchers also propose a future plan: if you can tune a two-part instrument, you can use the same logic to tune a chain of many parts, one by one. This could eventually allow scientists to build the long chains of quantum bits needed for advanced quantum computing tasks like "braiding" (twisting quantum states together).

In short: They taught a robot to tune a complex quantum instrument by letting it study theory, practice on a similar model, and then successfully tune a new, real device on its own, finding the perfect settings in under an hour.

Problem Statement
Contemporary mesoscopic quantum devices are reaching new limits in size and complexity, facilitating the experimental exploration of emergent quantum modes. However, this scalability introduces significant challenges in device tuning and control. While machine learning (ML) has shown promise in operating semiconductor quantum dots, a critical gap remains in developing foundational models that generalize across diverse experimental platforms and in demonstrating utility beyond simple charge stability diagrams with low parameter counts. Specifically, the autonomous tuning of complex emergent modes, such as Poor Man's Majorana zero modes (PMMs) in Kitaev chains, has not been previously achieved without relying on data from the specific experimental platform being tuned.

Methodology
The authors propose and demonstrate a cross-platform autonomous tuning protocol for a minimal Kitaev chain. The methodology relies on a transfer learning approach involving three distinct objects: a theoretical model, a two-dimensional electron gas (2DEG) device (Device B), and a one-dimensional nanowire device (Device A).

Theoretical Training: A Convolutional Neural Network (CNN) is initially trained on 6,000 simulated charge stability diagrams (CSDs) generated from an effective theory model of a three-site Kitaev chain. The model predicts the ratio of elastic co-tunneling (ECT) to crossed Andreev reflection (CAR) rates, labeled as $(t - \Delta)/(t + \Delta)$ , where $t = \Delta$ corresponds to the "sweet spot" where PMMs emerge.
Cross-Platform Validation and Retraining: The CNN is first validated on experimental data from Device B (2DEG). To improve accuracy and bridge the domain gap between theory and experiment, the model is retrained on a small subset (51 labeled CSDs) of experimental data from Device B.
Application to Target Device: The retrained model is applied to Device A (nanowire) without using any training data from Device A. The CNN predicts the $(t - \Delta)/(t + \Delta)$ ratio from measured local differential conductance ( $G_{LL} \cdot G_{RR}$ ).
Gradient Descent Optimization: These predictions serve as input for a gradient descent algorithm that adjusts the hybrid plunger gate voltage ( $V_{ABS}$ ) to minimize the cost function $f(V_{ABS}) = |(t - \Delta)/(t + \Delta)|$ , driving the system toward the sweet spot.

Key Contributions

Cross-Platform Transfer Learning: The work demonstrates that a tuning algorithm trained on theory and a different experimental platform (2DEG) can successfully tune a distinct device architecture (nanowire) without specific retraining on the target device's data.
Autonomous Emergent Mode Tuning: The protocol autonomously tunes a minimal Kitaev chain to the emergence of PMMs, moving beyond simple charge stabilization to the control of complex quasiparticle states.
Robustness to Magnetic Fields: The algorithm is tested under both zero and finite magnetic fields ( $B = 150$ mT), showing functionality in both regimes, though with varying convergence rates depending on interaction strength.

Results

Convergence Performance (Zero Field): In zero magnetic field, the algorithm successfully converges to the immediate vicinity of a sweet spot in 67.6% of attempts (within $\pm 1.5$ mV) and in 80.9% of cases (within $\pm 4.5$ mV). The typical time to find a sweet spot is 45 minutes or less.
Convergence Performance (Finite Field): Under a finite magnetic field ( $B = 150$ mT), convergence is less frequent due to weaker quantum dot interactions and lower conductance, which challenges the model trained on strongly interacting data. However, statistical analysis of multiple runs still identifies sweet spots, with 45.7% of runs converging within $\pm 1.62$ mV and 60% within $\pm 4.85$ mV of verified sweet spots.
Generalization: The CNN predictions on Device A closely match experimental values for small $(t - \Delta)/(t + \Delta)$ ratios, correctly identifying the sign change that indicates a sweet spot. The model also generalizes to a previously unseen finite-field dataset from a 2DEG device, confirming its ability to operate at finite magnetic fields provided the quantum dot coupling is sufficiently strong.

Significance
The authors position this work as a stepping stone toward the autonomous tuning of emergent modes in interacting systems and the development of foundational tuning ML models deployable across a range of experimental platforms. By successfully tuning a minimal Kitaev chain without platform-specific training data, the method addresses the scalability challenges of solid-state quantum devices. The paper concludes that this approach paves the way for tuning more complex, longer Kitaev chains required for topological quantum computing operations such as braiding and anyon fusion, proposing a protocol to extend the pairwise tuning algorithm to arbitrary chain lengths.

More like this