Interpretable ML Discovers Quantum Phenomena and Order Parameters

The arXiv paper 2604.16015 presents an end-to-end framework that uses interpretable machine learning to discover quantum phenomena and compact order parameters from unlabeled measurement data. The authors combine variational autoencoders for representation learning with symbolic regression to produce analytic descriptors. Results include applications to Rydberg-atom snapshots, classical shadows of the cluster Ising model, and hybrid fermionic data, and they report a novel corner-ordering pattern in Rydberg arrays. The authors release their code as the qdisc Python library.

The pipeline trains unsupervised generative encoders to learn low-dimensional latent representations that correlate with physical regimes. After representation learning, symbolic methods translate latent coordinates into closed-form analytical expressions that act as order parameters, rather than opaque features. Key components shown or implied in the paper are:

• •variational autoencoders for unsupervised compression and disentanglement of measurement distributions
• •classical shadows as a compact measurement representation for many-body states
• •symbolic regression to extract interpretable formulas from latent variables
• •evaluation on experimental and simulated datasets, including Rydberg snapshots and cluster Ising states

The approach emphasizes interpretability over black-box classification: learned descriptors are physically inspectable, compact, and experimentally measurable. The provided qdisc package bundles model training, symbolic extraction, and dataset hooks to reproduce the experiments.

This work sits at the intersection of representation learning, symbolic discovery, and quantum many-body physics. It advances the trend of replacing post-hoc explainability with intrinsic interpretability by designing pipelines that output analytic physical quantities. Prior studies used unsupervised networks to cluster phases; this paper closes the loop by producing explicit equations usable as order parameters. For experimental groups working with Rydberg arrays or near-term quantum devices, the methodology offers a practical path to find emergent structure without labeled phase diagrams. The open-source release accelerates reproducibility and adoption.

Validate the approach on larger system sizes, noisy experimental conditions, and circuit-based quantum hardware. Adoption will depend on robustness to measurement noise and how well symbolic descriptors generalize across parameter sweeps. The next steps are benchmarks against established order parameters and integration of qdisc into experimental control stacks.