Speaker
Description
Detecting phases of matter generally relies on identifying the correct order parameter - a task that remains notoriously difficult for unknown transitions and is traditionally guided by physical intuition. While neural networks have recently offered an alternative route by locating phase transitions without a priori physical knowledge, these approaches often remain black boxes. Moreover, they frequently struggle when confronted with realistic, noisy experimental data - the ultimate testbed for automated methods in physics.
Here, we bridge these perspectives by introducing TetrisCNN, a convolutional architecture with parallel branches of different filters designed to detect phases of matter and extract the symbolic form of order parameters directly from spin configurations. We demonstrate an automated discovery pipeline on raw projective measurements from Ising and XY models realised on Rydberg atom quantum simulators. First, a prediction-based method performs regression on experimental tuning parameters, pinpointing the phase transition via the peak derivative of these predictions. Second, using this discovered boundary as ground truth, we pivot to a supervised classification task, which corresponds to the order parameter search. Crucially, the network’s sparse latent space robustly selects the identical set of physical spin correlators across both tasks, allowing us to extract the symbolic form of the order parameter from classification results.
This data-driven approach demonstrates the ability to identify predictive signals that challenge standard theoretical assumptions. For the two-dimensional Ising model, the network selects simple magnetization as the most reliable phase indicator. For the XY model, moving beyond the traditional single-basis measurements standard in automated classification, a joint-basis analysis reveals that a single-body cross-basis correlator $\overline{S_{i}^x S_{i}^z}$ significantly improves prediction over purely $z$-basis data. While the complexity of the extracted features is systematically controlled by the bottleneck sparsity penalty, the selected correlators remain remarkably stable across random network initializations. This framework opens the way to integrating interpretable machine learning with quantum simulators to autonomously uncover new exotic phases of matter.