Unsupervised Machine Learning for Experimental Detection of Quantum-Many-Body Phase Transitions

Ron Ziv, David Wei, Antonio Rubio-Abadal, Daniel Adler, Anna Keselman, Eran Lustig, Ronen Talmon, Johannes Zeiher, Immanuel Bloch, Mordechai SegevarXiv 2025 · DOI · arXiv

TL;DR

Researchers have developed an unsupervised machine learning framework that detects phase transitions in quantum many-body systems directly from raw experimental "snapshots." By adapting a manifold learning technique called Diffusion Maps, the team successfully identified critical transition points in complex 2D systems—including Many-Body Localization—without needing prior theoretical models or predefined "order parameters."

What Problem Does This Paper Address?

Quantum Many-Body (QMB) systems are notoriously difficult to study because their complexity scales exponentially with the number of particles. While Richard Feynman’s vision of "quantum simulators" allows us to mimic these systems in the lab, we face a massive data problem: we can only see a tiny fraction of what’s happening.

In a typical experiment using a quantum gas microscope, researchers get "snapshots" of atom positions. However, these snapshots are often limited (e.g., only showing whether a site has an even or odd number of atoms). To find a phase transition (like the jump from a conductor to an insulator), physicists usually need an "order parameter"—a specific mathematical signature they know to look for. If the system is new or too complex for classical computers to simulate, we might not even know what that signature is. This paper tackles the challenge of finding these transitions blindly, using only the raw, partial data available from experiments.

Methods

The core of the approach is Diffusion Maps (DM), a type of unsupervised manifold learning. DM assumes that even if data looks like a high-dimensional mess, it actually lives on a much simpler, low-dimensional "manifold" governed by just a few effective degrees of freedom.

The authors introduced two critical innovations to make DM work for quantum data:

  1. Distribution-Aware Kernels: Standard ML looks at individual data points. However, quantum measurements are probabilistic. The authors modified the DM algorithm to measure the Mahalanobis-like distance between sets of snapshots. This allows the algorithm to compare the underlying probability distributions of different experimental settings rather than just comparing individual images.
  2. Wavelet Preprocessing: They applied a weighted wavelet transform to the snapshots. This step, related to the "Earth Mover's Distance," helps the algorithm focus on structural correlations in the atom distributions, which proved essential for detecting subtle phenomena like Many-Body Localization.

Crucially, the method is model-agnostic and fast. It requires no training labels and can process a full dataset in under a minute on a standard computer without a GPU.

Datasets

The researchers validated and applied their method across several regimes:

  • Numerical Benchmarks: 1D Transverse Field Ising Model (simulated via DMRG) and the 2D Classical Ising Model. These served as "ground truth" tests.
  • Experimental Mott-to-Superfluid Transition: Data from a 2D Bose-Hubbard system using square, triangular, and Lieb lattice geometries.
  • Experimental Many-Body Localization (MBL): 2D disordered systems. They tested both a "sharp edge" initial state (where a traditional "imbalance" metric exists) and a "uniform" initial state (where traditional metrics fail).

Key Insights and Ideas

  • The Emergent Order Parameter: The first latent coordinate ($\Phi_1$) extracted by the algorithm acts as an "emergent order parameter." When plotted against physical parameters (like magnetic field or disorder strength), it shows sharp "plateaus" and transitions that perfectly mirror physical phase changes.
  • Seeing the Invisible: In the MBL experiments with uniform initial states, the raw atom density images show no obvious signs of localization to the human eye. However, the ML algorithm successfully identified a clear transition point ($W_c = 9.5$), proving it can extract hidden collective phenomena that traditional heuristic measures miss.
  • Experimental Realism: The ML results for the Mott-to-Superfluid transition showed slight deviations from idealized Quantum Monte Carlo (QMC) simulations. The authors argue this isn't a failure of the ML, but rather evidence that the method is sensitive to real-world experimental factors like residual heating and trap inhomogeneities.

Key Results

  • Ising Model: The algorithm identified the critical point at $\lambda_c \approx 0.938$, very close to the theoretical value of 1.0, despite finite-size effects.
  • Mott-to-Superfluid: Successfully mapped the transition across three different lattice geometries, providing the first reported critical values for the Lieb lattice configuration.
  • MBL Detection: In the challenging 2D MBL case, the method identified a critical disorder strength of $W_c = 13.4$ for edge states and $W_c = 9.5$ for uniform states.
  • Efficiency: The framework is "on-the-fly" ready, executing in seconds/minutes, making it a candidate for real-time experimental monitoring.

Limitations and Future Work

The authors acknowledge a few key hurdles:

  • Transition vs. Crossover: While the method detects a "change in behavior," it cannot yet distinguish between a true thermodynamic phase transition and a very sharp crossover (a common debate in MBL research).
  • Physical Interpretation: The algorithm tells you where the transition is, but it doesn't automatically explain the physics of why it's happening. Connecting the latent coordinates back to specific physical symmetries remains an open theoretical question.
  • Adaptive Protocols: The next step is integrating this into "closed-loop" experiments, where the ML detects a transition in real-time and automatically tells the quantum simulator which parameters to explore next.

Summary updated

Abstract

Quantum many-body (QMB) systems are generally computationally hard: the computing resources necessary to simulate them exactly can often exceed the existing computation resources by orders of magnitude. For this reason, Richard Feynman proposed the concept of a quantum simulator: quantum systems engineered to obey a prescribed evolution equation and repeating the experiment multiple times. Experimentally, however, as we explain below, the vast majority of observables describing the system are inaccessible. Thus, while Feynman's idea addresses the problem of simulating quantum dynamics, it leaves unsolved the equally fundamental problem of inferring the underlying physics from the limited observables accessible in experiments. Indeed, many complex phenomena associated with QMB systems remain elusive. Perhaps, the most important example is identifying phase transitions in QMB systems when no simple order-parameter exists, which poses major challenges to this day. Complicating the problem further is the fact that, in most cases, it is impossible to learn from numerical simulations, as the underlying systems are often too large to be computable, and small QMB can show strong finite size effects, masking the presence of the transition. Here, we present an unsupervised machine learning approach to study QMB experiments, specifically aimed at detecting phase transitions and crossovers directly from raw experimental measurements. We demonstrate our methodology on systems undergoing Many-Body Localization cross-over and Mott-to-Superfluid phase-transition, showing that it reveals collective phenomena from the very partial experimental data and without any model-specific prior knowledge of the system. This approach offers a general and scalable route for data-driven discovery of emergent phenomena in complex quantum many-body systems.
BibTeX
@misc{ziv2025unsupervised,
  author = {Ron Ziv and David Wei and Antonio Rubio-Abadal and Daniel Adler and Anna Keselman and Eran Lustig and Ronen Talmon and Johannes Zeiher and Immanuel Bloch and Mordechai Segev},
  title = {{Unsupervised Machine Learning for Experimental Detection of Quantum-Many-Body Phase Transitions}},
  year = {2025},
  eprint = {2512.01091},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.01091},
  howpublished = {arXiv},
}

Loading related content...

Figure