arXiv:2609.20693
TetrisCNN for interpretable detection of phases of matter from experimental quantum simulator data
Kacper Cybiński, Björn van Zwol, James Enouen, Guillaume Bornet, Thierry Lahaye, Antoine Browaeys, Antoine Georges, Anna Dawid
Abstract
Detecting phases of matter in general relies on identifying the correct order parameter - a task that remains notoriously difficult for unknown transitions and traditionally is guided by physical intuition and educated guess. Neural networks have recently offered an alternative route by locating phase transitions in known models without any a priori physical knowledge. Yet these approaches remain black boxes and only identify phases without elucidating their properties. Moreover, they often struggle when confronted with realistic, noisy experimental data, which constitute the ultimate testbed for automated methods in physics. Here, we bridge these perspectives by introducing TetrisCNN, a convolutional architecture with parallel branches of differently shaped filters, reminiscent of Tetris blocks, that learns sparse, interpretable latent representations directly in terms of spin correlators. Applied to experimental snapshots of two-dimensional Ising and XY quantum simulators measured in multiple bases, the network not only detects phase transitions and crossovers but also expresses its latent representation and decision boundaries as symbolic formulas built from experimentally measurable spin correlators. This framework opens the way to integrating interpretable neural networks with quantum simulators to uncover and understand new phases of matter.
AI Summary
The paper asks how neural-network analysis of experimental quantum-simulator snapshots can expose the observables behind its predictions. TetrisCNN uses parallel convolutional branches with different spatial supports, global average pooling, and size-dependent sparsity penalties. A Boolean-Fourier expansion expresses each branch as a linear combination of spin correlators within its receptive field; regression extracts these combinations, and polynomial surrogates describe the classifier’s decision boundary. The demonstrations use existing $8\times8$ Ising and $6\times7$ XY Rydberg-array datasets. Five established unsupervised methods first estimate changes along experimental sweeps; these estimates then define labels for supervised TetrisCNN training. Validation accuracies are $99.14\%$ for Ising and $93.54\%$ for XY. For Ising, only uniform $Z$ magnetization survives: the detected change near $1.1\,\mu\mathrm{s}$ is a polarization crossover, not the later antiferromagnetic transition, which none of the tested detection methods identifies. For XY, three branches predominantly encode $X$ magnetization and nearest-neighbor and diagonal $Z$ correlators. A quadratic boundary in branch coordinates reproduces $99.84\%$ of network predictions; its approximate correlator-space version reaches $98.30\%$. The squared $X$-magnetization contribution relates to the established in-plane order parameter. Repeated training and snapshot re-pairing test descriptor stability. The selected observables are task-dependent discriminators, not necessarily order parameters, and may reflect experimental artifacts. Interpretation scales exponentially with receptive-field size and currently becomes impractical for long-range correlations. The datasets are finite, noisy, nonequilibrium sweeps; no independent test set is used, and the reported validation accuracies are not intended as generalization benchmarks.
Error found: The rotation-averaging operation specified in the equations can generate correlators outside the subpatterns assigned to a branch, invalidating the stated branch-support guarantee unless rotations are instead averaged after the nonlinear branch computation.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 34 pages, 25 figures