Researchers have devised a neural network capable of classifying quantum phases, specifically, distinguishing between ‘trivial,’ ‘long-range entangled,’ and ‘symmetry-protected topological’ states, directly from raw measurement outcomes. This new approach bypasses the need to reconstruct the complex quantum state itself, a computationally intensive process that previously limited the study of these transitions to small systems.
The system’s success is signaled by a “sharp convergence in the classifier’s accuracy,” indicating the system has reached a steady-state phase without relying on the exponentially unfeasible process of Hui Yu and Jiangping Hu of the Chinese Academy of Sciences detail this robust, post-selection-free method for experimentally probing measurement-induced phases.
Neural Network Classifies Measurement-Induced Area-Law Phases
This network classifies phases by directly analyzing raw measurement outcomes, sidestepping the need for multiple identical quantum trajectories traditionally required for observation of measurement-induced phase transitions. The architecture combines a convolutional neural network with an attention mechanism, a specific design choice intended to capture correlations between quantum trajectories. The team’s approach addresses a longstanding challenge in the field: accurately identifying multiple coexisting area-law phases, which are characterized by entanglement patterns that scale with the boundary of the quantum system, rather than its volume.
Previous methods often focused on differentiating only two phases, typically volume-law and area-law regimes, but distinguishing between more subtle area-law states remained elusive. By using machine learning, the researchers created a system that operates directly on classical data records, eliminating the need for complex calculations of nonlinear observables like mutual information or topological entanglement entropy.
This simplification is critical, as calculating these quantities becomes exponentially more difficult as the size of the quantum system increases. The neural network’s performance was systematically evaluated across a range of parameters, including sample size, input data, spatial and temporal constraints and system size scalability. The system’s architecture begins by reshaping raw measurement records, specifically, X, ZZ and ZXZ measurements, to align with the circuit’s brickwork periodicity.
These inputs are then processed by three parallel convolutional neural network branches, each independently extracting features. These features are subsequently combined, projected through a linear layer, and compressed using a temporal readout before being fed into a pooling attention module. This module aggregates representations from multiple trajectories into a single vector, which is then mapped to a probability distribution over the three target phases, trivial, long-range entangled and symmetry-protected topological, using a softmax layer.
The researchers detail that their weak measurement technique is mathematically defined by Kraus operators, constructed using a mixing angle dependent on a parameter . After each measurement, the wavefunction is explicitly normalized to ensure a proper post-measurement state. This normalization process, detailed in the supporting materials, is a key component of the overall methodology. The ability to classify phases directly from raw measurement outcomes represents a step toward practical experimental verification of these complex quantum states, offering a robust and scalable alternative to traditional methods.
Measurement-Only Circuits Define Trivial, LR, and SPT Phases
This bypasses a historically intensive process, offering a pathway toward practical experimental verification of these elusive phases of matter. The new method uses a neural network architecture combining convolutional and attention mechanisms to achieve this post-selection-free phase classification. The core of the advance lies in the network’s ability to discern subtle differences in measurement records generated by systems subjected to competing projective measurements. These measurements, involving the application of X, ZZ, and ZXZ gates, drive transitions between area-law entangled states, but traditionally identifying these transitions demanded calculating complex entanglement quantities.
The team demonstrated that integrating data from all three measurement channels is critical for achieving high classification accuracy, with performance reaching over 80 percent for test points located away from phase boundaries. The researchers systematically investigated how various factors influence the network’s performance.
The number of sampled trajectories, the extent of measurement data used, the duration of the final evolution, and the spatial size of the input all play a role in classification accuracy. The team found that a network trained on smaller systems can generalize its learned features to larger system sizes without requiring additional training, suggesting a significant step toward applying the method to more complex and realistic quantum systems. The researchers detail that their method is robust and offers a practical pathway for experimentally probing measurement-induced phases.
Post-Selection Problem Limits Experimental Verification
The conventional approach to verifying these transitions relies on calculating entanglement quantities, a process demanding exponentially increasing computational resources as system size grows. This new methodology addresses a significant hurdle in experimentally observing measurement-induced phase transitions, by using the power of modern machine learning. The system’s performance was evaluated using circuits of up to 12 qubits, with the network trained to recognize the three distinct area-law phases generated by varying the strengths of different types of quantum measurements.
The architecture’s design is important to its success, processing raw measurement records, the outcomes of each quantum measurement, directly without requiring intermediate calculations of complex quantum states. Three parallel convolutional neural network branches independently analyze the data from each measurement type, X, ZZ and ZXZ, before combining the extracted features.
This parallel processing is then followed by spatial global average pooling and a temporal readout, ultimately feeding into a pooling attention module that aggregates information from multiple quantum trajectories. The final output is a probability distribution indicating the likelihood of the system being in each of the three phases. The circuit model employed in the study uses a one-dimensional brickwork architecture with open boundary conditions, allowing for a systematic investigation of the phase transitions.
The system is initialized in a specific state and evolved over time using generalized weak measurements, ensuring that the network receives a clear signal from the underlying entanglement structure. The strength of each measurement type is controlled by a parameter, γ, subject to the constraint that the sum of all γ values equals one. This parameter allows the researchers to tune the system and explore the boundaries between the different phases. Further analysis revealed the importance of the attention mechanism within the neural network.
Scalability of Neural Networks to Larger Quantum Systems
The ability of a neural network to maintain classification accuracy even when applied to quantum systems twice its original training size demonstrates a key advance in probing complex quantum phenomena. Unlike prior methods requiring reconstruction of the full quantum state, a computationally intensive process, this approach bypasses that step entirely. Partial measurement records, drawn from a restricted spatial region, proved surprisingly effective for accurate phase identification, indicating that information at smaller scales is sufficient for robust classification.
Further analysis revealed the network’s capacity to generalize beyond the training parameters. This resilience suggests the network extracts universal features of the three phases, rather than being limited by the specific parameters of the initial training run. The researchers also investigated the impact of input data on classification performance.
This is particularly important for experimental implementations, where acquiring complete measurement records can be challenging and time-consuming. The ability to achieve high accuracy with partial data significantly reduces the experimental overhead and makes it more feasible to study these phases in larger systems. The study’s implications extend beyond the specific phases investigated. The demonstrated ability to classify quantum states directly from measurement outcomes could be applied to a wider range of quantum systems and phenomena.
Funding for this research was provided by the Innovation Program for Quantum Science and Technology (2024ZD0301700) and the National Natural Science Foundation of China (No. 12574546). Hui Yu, Jiangping Hu, and Shi-Xin Zhang jointly conceived the project and designed the research. Hui Yu performed the numerical calculations and data analysis under the supervision of Shi-Xin Zhang and Jiangping Hu. “All authors wrote the manuscript and discussed the results,” according to the published paper.
This new framework offers a robust and scalable solution for analyzing complex quantum systems, potentially leading to new understanding of the behavior of matter at its most fundamental level.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
