⚛️ quantum physics

Machine-Learning-Empowered Quantum Sensing of the Plaquette Phase in a Three-Level Delta System

This paper proposes a supervised machine-learning approach using a multi-layer perceptron to accurately estimate the gauge-invariant plaquette phase in three-level $\Delta$ systems by analyzing STIRAP population transfer efficiencies, thereby enabling effective phase identification for quantum sensing applications.

Original authors: Lorenzo Vitale, Shreyasi Mukherjee, Dario Fasone, Enrico Martello, Elisabetta Paladino, Luigi Giannelli, Giuseppe Falci

Published 2026-07-17

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Lorenzo Vitale, Shreyasi Mukherjee, Dario Fasone, Enrico Martello, Elisabetta Paladino, Luigi Giannelli, Giuseppe Falci

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 listen to a secret message hidden inside a complex song. In the world of quantum physics, scientists often try to control tiny particles, like atoms or artificial atoms, using lasers. These particles can exist in multiple states at once, a spooky feature called "superposition." One of the most famous tricks in this field is called "coherent population trapping." Think of it like a magic trick where a particle is so perfectly balanced between two states that it refuses to jump to a third, "lossy" state that would ruin the experiment. It's like a tightrope walker who, thanks to perfect balance, never falls off the rope.

Usually, scientists love this perfect balance because it keeps the system stable. But what if that perfect balance was actually a clue? What if the reason the tightrope walker wobbled wasn't a mistake, but because the wind (an invisible force) was blowing in a specific direction? This is the question at the heart of a new study. The researchers are looking at a specific setup where three energy levels form a triangle, creating a hidden "phase" or angle that acts like a secret compass direction. This phase is invisible and can't be measured directly, but it changes how the particles dance. The big question is: Can we figure out this secret angle just by watching how the particles move, even when they aren't moving perfectly?

This paper proposes a clever solution: use a computer brain, specifically a type of artificial intelligence called a machine learning model, to act as a detective. The researchers simulated a three-level quantum system shaped like a triangle (a "delta" system). In this system, there is a hidden "plaquette phase"—a gauge-invariant quantity that arises because the three levels are connected in a loop. Normally, scientists use a technique called STIRAP (Stimulated Raman Adiabatic Passage) to move particles from one state to another efficiently. This technique relies on that perfect "trapped state" we mentioned earlier. However, the researchers found that when this hidden plaquette phase is present, it breaks the perfect trapping. The particles leak out a little bit, and the efficiency of the transfer drops.

Instead of seeing this leakage as a failure, the authors suggest it's a treasure trove of information. They created a dataset by running thousands of computer simulations where they changed the hidden phase and measured how much the transfer efficiency dropped under different conditions. They then fed this data into a multi-layer perceptron (MLP), a type of neural network. Think of the MLP as a student who has never seen the secret phase but has been shown millions of examples of "how much the particle leaked" for different angles. The goal was to see if the computer could learn the pattern and guess the secret angle just by looking at the leakage.

The results of these simulations were quite promising. The neural network, trained on the simulated data, learned to map the transfer efficiencies to the hidden phase with high accuracy. The researchers tested the model on new data it had never seen before, and it successfully reconstructed the phase across the entire range of possibilities. The "error" in the prediction was very small, suggesting that the combination of the specific driving conditions and the machine learning model is a powerful way to "sense" this invisible phase.

Crucially, the paper emphasizes that this is a simulation-based proof of concept. The authors show that the method can work in a controlled digital environment, but they haven't yet tested it on a real physical quantum device in a lab. They argue that this approach turns a problem (the breakdown of perfect trapping) into a resource (a sensor for the phase). They suggest that this technique could be applied to real-world platforms like superconducting circuits or quantum dots, where such triangular systems can be engineered. However, they also note that future work is needed to test how well this holds up against real-world noise and to see if fewer measurements are needed to get the same result.

In short, the paper demonstrates that by combining a specific quantum control protocol with a machine learning algorithm, we can potentially "read" an invisible quantum phase by observing how the system's performance degrades. It's like teaching a computer to recognize the shape of the wind by watching how a leaf falls, rather than trying to measure the wind directly. This opens up a new perspective where imperfections in quantum control aren't just bugs to be fixed, but features to be exploited for sensing the hidden geometry of the quantum world.

Technical Summary: Machine-Learning-Empowered Quantum Sensing of the Plaquette Phase

Problem Statement
The paper addresses the challenge of sensing the "plaquette phase," a gauge-invariant quantity arising in closed-loop three-level $\Delta$ -type quantum systems. Unlike open $\Lambda$ -type systems, $\Delta$ -systems possess a triangular coupling structure where a residual complex phase ( $\phi$ ) cannot be eliminated by local unitary transformations. This phase profoundly influences system dynamics, specifically by breaking the conditions for perfect Coherent Population Trapping (CPT). While this breakdown degrades the performance of standard population transfer protocols like Stimulated Raman Adiabatic Passage (STIRAP), it creates a non-trivial, phase-dependent leakage that encodes information about the otherwise inaccessible plaquette phase. The core problem is formulated as a non-linear inverse problem: reconstructing the phase $\phi$ from experimentally accessible observables (transfer efficiencies) given the highly non-trivial and generally non-invertible mapping between the phase and these observables.

Methodology
The authors propose a data-driven framework combining coherent quantum control with supervised machine learning (ML) to solve this inverse problem.

System Model: The study utilizes a three-level $\Delta$ -system described by a Hamiltonian with three mutually coupled states. The couplings are parameterized to include a gauge-invariant plaquette phase $\phi$ . The dynamics are governed by the breakdown of the dark state (trapped state) when $\phi \neq 0, \pm\pi$ , leading to phase-dependent leakage into the intermediate level.
Sensing Protocol: The authors employ a STIRAP-like protocol to transfer population from state $|0\rangle$ to $|1\rangle$ . Instead of attempting to engineer a complex detuning to perfectly restore trapping for unknown phases, they fix the detuning to the value corresponding to the zero-phase case. Under this operational condition, the transfer efficiency ( $\eta$ ) becomes a function of the plaquette phase $\phi$ .
Data Generation: A synthetic dataset is generated by numerically solving the time-dependent Schrödinger equation for 2,000 uniformly spaced values of $\phi \in [-\pi, \pi]$ . The input features for the machine learning model consist of an 8-dimensional vector of transfer efficiencies. These efficiencies are measured under varying driving conditions (different ratios of pump and Stokes pulse amplitudes) and detuning protocols (zero detuning, symmetric trapping detuning, and matched driving-detuning conditions).
Machine Learning Architecture: A Multi-Layer Perceptron (MLP) is trained in a supervised learning framework.
- Input: The 8-dimensional vector of STIRAP efficiencies.
- Output: To handle the periodic nature of the phase ( $2\pi$ ), the target is encoded as a 2D vector $(\cos \phi, \sin \phi)$ on the unit circle, avoiding boundary discontinuities.
- Training: The network uses a Mean Squared Error (MSE) loss function and is optimized via the Adam algorithm. The dataset is split into training (60%), validation (20%), and test (20%) sets.

Key Contributions

Reframing Imperfections as Resources: The work demonstrates that the breakdown of coherent population trapping, typically viewed as a limitation in control protocols, can be harnessed as a resource for quantum sensing.
ML-Assisted Phase Reconstruction: The authors successfully demonstrate that a neural network can accurately reconstruct the plaquette phase from a limited set of experimentally accessible observables (transfer efficiencies) without requiring direct measurement of the phase itself.
Operational Protocol: They define a practical sensing protocol that does not require prior knowledge of the phase to set the detuning, relying instead on the phase sensitivity of the efficiency under fixed, standard driving conditions.

Results
The trained MLP model achieves high accuracy in reconstructing the plaquette phase.

Convergence: The cost function converges to small values ( $10^{-3}$ to $10^{-4}$ ) with stable training curves, showing no signs of overfitting between training and validation sets.
Prediction Accuracy: On the test set (unseen data), the predicted phase values align closely with the true target values along the identity line. The prediction error remains consistently small across the entire dataset.
Generalization: The results indicate that the selected 8-dimensional input features encode sufficient information to resolve the phase over the full interval $[-\pi, \pi]$ , and the model generalizes effectively beyond the training data.

Significance and Claims
The paper claims that this approach highlights how the combination of coherent control and machine learning enables effective phase identification in closed-loop quantum systems. The authors posit that this method opens new perspectives for quantum technologies, specifically for quantum sensing applications involving synthetic gauge fields.

The work emphasizes a broader paradigm shift: imperfections in coherent control protocols can be repurposed as flexible and scalable tools for quantum sensing. By integrating physically motivated control schemes with data-driven inference, the authors provide a framework for addressing non-linear inverse problems in quantum systems. The proposed approach is noted as being compatible with various experimental platforms capable of engineering $\Delta$ -type configurations, including superconducting circuits, nitrogen-vacancy centers, and quantum dots. The authors modestly suggest that future work could focus on reducing the number of required input features and exploring alternative driving schemes to further enhance experimental feasibility.

Technical Summary: Machine-Learning-Empowered Quantum Sensing of the Plaquette Phase

More like this