🔬 mesoscale physics

Machine-learned tuning of artificial Kitaev chains from tunneling-spectroscopy measurements

This paper demonstrates that a machine-learning algorithm using tunneling spectroscopy features from sensor dots can reliably tune artificial Kitaev chains to high-quality Majorana sweet spots by balancing crossed Andreev reflection and elastic cotunneling, offering a scalable approach for achieving topological protection in longer chains.

Original authors: Jacob Benestad, Athanasios Tsintzis, Rubén Seoane Souto, Martin Leijnse, Evert van Nieuwenburg, Jeroen Danon

Published 2026-02-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jacob Benestad, Athanasios Tsintzis, Rubén Seoane Souto, Martin Leijnse, Evert van Nieuwenburg, Jeroen Danon

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 build a very delicate, invisible bridge made of light. This bridge isn't for people; it's for a special kind of particle called a Majorana. In the world of quantum physics, these particles are like "ghosts" that are incredibly useful for building super-fast, unbreakable computers. However, they are also extremely shy and fragile. If you touch them too hard or set up the bridge wrong, they disappear or turn into something useless.

The problem is that building these bridges (called Kitaev chains) is like trying to tune a radio with a million knobs, all at once, while blindfolded. You have to adjust the voltage on tiny quantum dots (tiny islands where electrons live) to a "sweet spot" where the Majorana particles feel safe. If you miss the spot by a tiny fraction, the whole thing fails.

Traditionally, scientists tried to tune these knobs one by one, like a human trying to find a radio station by turning one dial, listening, then turning another. But as the chain gets longer, this becomes impossible. There are too many variables, and the "noise" in the system makes it hard to know if you're getting closer or further away.

The Solution: A Smart, Automated Tuner

This paper introduces a new way to tune these quantum bridges using Machine Learning. Think of it as hiring a super-smart, tireless robot tuner instead of a human.

Here is how the robot works, using a few simple analogies:

1. The "Sensor Dots" are like Ears
The researchers added two extra "sensor dots" to the ends of their quantum chain. Imagine these as ears listening to the bridge. When the robot changes the voltage knobs, the ears listen to the "sound" (tunneling spectroscopy) the bridge makes.

The Goal: The robot wants to find a setting where the "sound" is perfectly silent (zero energy). This silence means the Majorana particles are happy and stable.

2. The "Loss Function" is a Scorecard
The robot doesn't know exactly where the sweet spot is. Instead, it has a scorecard called a Loss Function.

If the robot turns a knob and the "sound" gets louder (more energy), the score goes up (bad).
If the robot turns a knob and the "sound" gets quieter (closer to zero), the score goes down (good).
The robot's only job is to get the score as close to zero as possible.

3. The "CMA-ES" Algorithm is a Smart Hiker
The robot uses a specific strategy called CMA-ES. Imagine a hiker trying to find the deepest valley in a foggy mountain range.

The Mistake: A normal hiker might take a step, see if they are lower, and keep going. But if the terrain is bumpy, they might get stuck in a small dip (a local minimum) and think they are at the bottom.
The Smart Hiker: This robot doesn't just take one step. It throws a whole group of "explorers" (a cloud of possibilities) out into the fog at once. It listens to where they land. If a group of explorers finds a lower spot, the robot moves the entire group toward that spot and narrows their search. It learns from the collective experience of the group rather than just one path.

What They Discovered

The team tested this robot on two types of bridges:

A Short Bridge (2 sites): They knew exactly where the sweet spot was. The robot found it almost every time, proving it works.
A Longer Bridge (3 sites): They didn't know where the sweet spots were. The robot still found them! It discovered that sometimes the bridge doesn't need to be perfectly symmetrical to work; it just needs the right balance of forces.

Why This Matters

The biggest breakthrough here is automation.

Before: Tuning a quantum chain was like trying to solve a Rubik's cube while wearing oven mitts. It was slow, manual, and prone to error.
Now: The robot can tune all the knobs simultaneously. It looks at the global "sound" of the whole system and adjusts everything at once to find the perfect balance.

This is a crucial step toward building longer quantum chains. Longer chains are needed to protect the Majorana particles from the outside world (topological protection), which is the key to making them useful for real quantum computers.

In a nutshell: The researchers taught a computer to "listen" to a quantum system and automatically twist the right knobs to make the system sing in perfect harmony, paving the way for more stable and powerful quantum computers.

1. Problem Statement

The realization of Majorana Bound States (MBSs) in artificial Kitaev chains is a central goal in condensed matter physics for topological quantum computing. While "Poor Man's Majoranas" (PMMs) have been demonstrated in minimal two-site chains, scaling these systems to longer chains (required for topological protection) faces significant challenges:

Parameter Complexity: Tuning a chain requires balancing Crossed Andreev Reflection (CAR) and Elastic Cotunneling (ECT) processes across multiple links simultaneously.
Imperfect Information: Experimental imperfections (material disorder, varying gate lever arms) make precise theoretical modeling difficult.
Inefficiency of Manual Tuning: Current methods rely on sequential, pairwise manual tuning. This is time-consuming and prone to error, as tuning one pair can disrupt previously tuned sections, especially in chains with strong tunnel couplings where multi-dot renormalization effects occur.
Need for Automation: A global metric is needed to tune all on-site potentials simultaneously without relying on derivative-based optimization, which is often unsuitable due to the non-smooth nature of experimental data and the presence of multiple local minima.

2. Methodology

The authors propose an automated tuning framework using the Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES) algorithm, guided by a specific loss function derived from tunneling spectroscopy.

A. System Model

Architecture: A linear array of quantum dots where every other dot is proximitized by a superconductor (creating Andreev bound states).
Sensors: Extra "sensor dots" are placed at both ends of the chain (Left/Right).
Hamiltonian: The system is modeled using a tight-binding Hamiltonian ( $H$ $H$ ) that includes:
- On-site energies ( $\epsilon$ ) and Zeeman splitting ( $E_Z$ ).
- Coulomb interactions ( $U$ ) on normal dots.
- Proximity-induced pairing ( $\Delta$ ) on superconducting dots.
- Spin-conserving and spin-non-conserving hopping ( $t, t^{SO}$ ).
Simulation Cases: The authors simulate 2-site (minimal) and 3-site artificial Kitaev chains.

B. The Tuning Algorithm (CMA-ES)

Optimization Strategy: CMA-ES is a derivative-free stochastic optimization method. It samples parameter configurations (gate voltages) from a multivariate normal distribution, evaluates their "fitness," and updates the distribution's mean and covariance matrix to converge toward optimal solutions.
Why CMA-ES? It handles non-smooth objective functions and multiple local minima better than gradient-based methods, making it robust against the noise and broadening inherent in real tunneling spectroscopy.

C. The Loss Function

The core innovation is a loss function ( $L$ ) based on local tunneling spectroscopy:

Metric: The algorithm measures the energy splitting between the lowest even and odd parity ground states ( $\delta E_{eo}$ ) as a function of the sensor dot potential ( $\epsilon_s$ ).
Objective: The goal is to find parameters where this splitting vanishes ( $\delta E_{eo} \approx 0$ ) while maximizing the Majorana Polarization (MP) ( $M$ ), which quantifies the quality of the MBS.
Loss Definition:
$L(\epsilon) = f(\epsilon) + \alpha \left\| \frac{\epsilon}{\Delta} \right\|^2$
- $f(\epsilon)$ : The maximum absolute energy splitting over a sweep of the sensor dot potential, averaged for left and right sensors.
- L2 Regularization: The term $\alpha \|\epsilon\|^2$ penalizes large gate voltages to keep the system within a physically reasonable range and bias the search toward the origin if multiple solutions exist.

3. Key Contributions

Global Tuning Metric: Demonstrated that local tunneling spectroscopy via end-sensor dots provides a reliable global metric for tuning entire chains, eliminating the need for sequential pairwise adjustments.
Algorithm Validation: Successfully applied CMA-ES to navigate high-dimensional parameter spaces to find "sweet spots" (points of zero splitting and high MP).
Scalability Proof: Showed that the method works not only for known 2-site configurations but also for 3-site chains where sweet spot locations are unknown a priori.

4. Results

A. Two-Site Chain (Benchmark)

Setup: 50 simulation runs targeting a known 2-site sweet spot.
Outcome: The algorithm consistently converged to the known sweet spots.
- Left Sweet Spot: Found in 41/50 runs. Final parameters deviated by Euclidean distance $\approx 0.028\Delta$ from the theoretical optimum. Resulted in $\delta E_{eo} \approx 2.7 \times 10^{-3}\Delta$ and $M \approx 0.967$ .
- Right Sweet Spot: Found in 9/50 runs. Similar high quality ( $M \approx 0.977$ ).
Convergence: Typically achieved within 80–100 generations.

B. Three-Site Chain (Unknown Landscape)

Setup: 50 simulation runs on a 3-site chain (5 dots total) with L2 regularization ( $\alpha = 0.02$ ).
Outcome:
- Convergence: 42/50 runs converged to a specific sweet spot with Majorana Polarization $M = 0.998$ and splitting $\delta E_{eo} < 10^{-3}\Delta$ .
- Symmetry Breaking: 8 runs converged to non-symmetric configurations that still yielded high MP and low splitting. This highlights that in longer chains, multiple distinct sweet spots (symmetric or asymmetric) can exist where CAR/ECT balance is achieved.
- Quality: The 3-site chain achieved higher MP values than the 2-site chain, confirming that adding sites improves the quality of the emergent modes.
Convergence: Typically required 140–200 generations due to the increased dimensionality.

5. Significance and Conclusion

Automation of Topological Tuning: This work provides a viable pathway to automate the tuning of complex quantum dot arrays, a critical bottleneck for realizing topologically protected qubits.
Robustness: The use of a derivative-free algorithm combined with a simple spectroscopic metric makes the approach robust against experimental noise and disorder.
Scalability: The success in the 3-site case suggests the method is scalable to longer chains, which are necessary for true topological protection.
Experimental Feasibility: The proposed metric (tunneling spectroscopy with sensor dots) is experimentally accessible, requiring only standard differential conductance measurements rather than complex interferometry setups.

In summary, the paper demonstrates that machine learning, specifically CMA-ES, can effectively navigate the complex parameter space of artificial Kitaev chains to locate high-quality Majorana sweet spots using simple, local spectroscopic data, paving the way for the automated fabrication of topological quantum devices.