Unsupervised Machine Learning Detects Quantum-Many-Body Phase Transitions Despite Limited Observables

Understanding the behaviour of complex quantum many-body systems presents a significant challenge, as accurately simulating these systems often demands computational resources far exceeding current capabilities. Ron Ziv, David Wei, and Antonio Rubio-Abadal, alongside colleagues at Technion and the Max-Planck-Institut für Quantenoptik, address this problem by developing a novel machine learning technique to identify phase transitions directly from experimental data. Their approach overcomes the limitations of relying on complete system knowledge or extensive simulations, which are often impractical due to system size and complexity. The team demonstrates the effectiveness of their unsupervised method by successfully detecting transitions in systems undergoing Many-Body Localization and Mott-to-Superfluid crossovers, revealing collective phenomena from only partial experimental measurements and without any pre-existing assumptions about the underlying physics. This work establishes a scalable, data-driven pathway for discovering emergent behaviour in a wide range of complex quantum systems.

Machine Learning Reveals Many-Body Localization Phase Transition

This research details a powerful new approach to understanding many-body localization (MBL), a fascinating quantum phenomenon where disorder can prevent a system from reaching thermal equilibrium. Scientists are using machine learning (ML) techniques to analyze data from experiments simulating quantum systems with cold atoms, revealing the transition to an MBL state and pinpointing the precise conditions where it occurs. This innovative method potentially bypasses the need for complex theoretical calculations. The team utilizes quantum simulators, creating systems with individual atoms arranged in a lattice using lasers and capturing snapshots of the atomic configurations to generate a large dataset.

These snapshots are then analyzed using diffusion maps, a machine learning technique that reduces data complexity while preserving underlying structure, allowing scientists to identify different phases of matter. Neural networks are also employed to classify the data and further refine the identification of the phase transition. The results demonstrate that machine learning can successfully identify the MBL phase and distinguish it from the thermal phase, pinpointing the critical point where the transition occurs. Researchers are also exploring the connection between MBL and brane parity order, a nonlocal property of the system, validating their findings by comparing them to theoretical predictions and strengthening our understanding of MBL.

Inferring Quantum Phase Transitions From Snapshots

Researchers have developed a new machine learning approach that can detect phase transitions in quantum systems directly from experimental measurements, overcoming limitations inherent in both simulating these complex systems and directly observing their behavior. Recognizing that simulations are computationally expensive and experiments often provide incomplete information, the team focused on inferring underlying physics from limited data, analyzing snapshots of atom density captured by quantum gas microscopes. The methodology addresses the challenge of identifying phase transitions when experiments only yield bulk quantities, which often fail to reveal a transition in quantum systems. The team’s approach circumvents the need for prior knowledge of the system or specific indicators of the transition, instead relying on machine learning algorithms to identify emergent patterns directly from the experimental data.

Recent experimental protocols enhance data acquisition by extracting additional observables, increasing the information content of each snapshot. The study demonstrates the methodology on systems undergoing a Many-Body Localization crossover and exhibiting a Mott-to-Superfluid phase transition. The results show that the unsupervised machine learning approach successfully reveals collective phenomena even with limited experimental data, offering a scalable route for data-driven discovery in complex quantum many-body systems.

Learning Quantum Order Directly From Data

This research presents a new machine learning approach for analyzing complex many-body quantum systems directly from experimental measurements, bypassing the need for detailed prior knowledge or computationally intensive simulations. Scientists successfully demonstrated the ability of this method to detect phase transitions and crossovers, even in systems too large for conventional computational techniques. The core breakthrough lies in a modified dimensionality reduction technique that learns an embedding, mapping experimental data, sets of single-shot configurational snapshots, into a lower-dimensional space. This allows the identification of emergent order parameters directly from the raw data, revealing underlying physics inaccessible through traditional averaging methods.

Initial validation involved applying the methodology to a simulated one-dimensional transverse field Ising model, a system known to undergo a phase transition, observing a sharp change in trend at a critical field value closely matching theoretical predictions. Further testing on a two-dimensional classical Ising system confirmed the methodology’s ability to accurately identify phase transitions and critical temperatures. Applying the technique to experimental data from two-dimensional quantum many-body systems, specifically investigating the Mott-to-superfluid phase transition and many-body localization scenarios, revealed the ability to identify transitions even in these complex systems, which are computationally intractable for simulations. The team’s approach offers a scalable route for data-driven discovery of emergent phenomena in complex many-body systems.

Machine Learning Reveals Quantum Phase Transitions

This research presents a novel machine learning approach for studying complex many-body quantum systems, directly addressing the challenge of inferring underlying physics from limited experimental data. The team developed a modified Diffusion Maps algorithm, adapted to handle the statistical nature of quantum measurements, enabling the identification of phase transitions and crossovers without prior knowledge of the system’s Hamiltonian or specific observables. Demonstrations on systems undergoing Many-Body Localization and Mott-to-Superfluid transitions reveal the method’s ability to detect collective phenomena, even when traditional approaches fail. The key achievement lies in the framework’s measurement-agnostic and state-agnostic nature, operating directly on raw experimental data and requiring minimal computational resources, typically executing in under a minute without GPU acceleration. This computational efficiency, combined with broad applicability across quantum platforms, offers a practical path towards scalable, model-independent analysis of quantum many-body systems and the potential discovery of new phenomena. Future work envisions integrating this framework into adaptive experimental protocols, enabling real-time detection of transitions and efficient mapping of complex quantum dynamics, ultimately guiding hypothesis formation in previously unknown models.