Interactive Explainer

Hybrid Quantum-Classical Neural Networks: Recognizing Quantum Phases

How a parameterized quantum circuit feeding a classical neural network can tell which phase of matter a many-body ground state belongs to. Based on recent arXiv work on quantum phase recognition shared by AndreasAtETH.

The core idea

Classifying phases of quantum matter is hard for purely classical machine learning because ground states of many-body systems live in an exponentially large Hilbert space. A hybrid architecture splits the job: a quantum feature circuit processes copies of the input state and produces measurement outcomes, and a classical neural network turns those measurement statistics into phase labels.

The pipeline mirrors the figure in the paper: (a) an input state of qubits from a spin chain, (b) shallow parameterized quantum layers, (c) single-qubit measurements yielding bitstrings, and (d) a small classical network head that outputs phase probabilities.

quantumclassicalmeasurement

The physical model

A standard benchmark is a spin-1/2 cluster chain with competing couplings. Its ground state phase depends on two parameters: a nearest-neighbor Ising coupling J1 and a next-nearest-neighbor coupling J2, both relative to the cluster term. Three phases appear: a symmetry-protected topological (SPT) cluster phase, a ferromagnetic phase, and an antiferromagnetic phase.

Drag the sliders below. The simulator prepares a proxy ground state, runs the quantum feature circuit with a finite number of measurement shots, and the classical head outputs phase probabilities. Fewer shots means noisier statistics, exactly the tradeoff a real device faces.

Predicted phase-
String order proxy-
Magnetization proxy-
Shot noise sigma-

This browser simulator uses analytic order-parameter proxies plus binomial shot noise rather than a full state-vector simulation, so it runs instantly while reproducing the qualitative behavior reported in the paper: sharp classification deep inside phases, uncertainty near boundaries, and degradation at low shot counts.

Explore the phase diagram

Each point in the plane below is a Hamiltonian. Colors show which phase the hybrid network assigns. Click or tap anywhere on the diagram to move the sliders there and re-run the classifier. White dashed lines mark the approximate true phase boundaries; the crosshair is your current point.

Keyboard: focus the diagram and use arrow keys to move the crosshair. Teal is the SPT cluster phase, violet is ferromagnetic, amber is antiferromagnetic.

Why hybrid beats either alone

Quantum part: feature extraction

The shallow circuit implements a coarse-graining map related to quantum error correction. It amplifies nonlocal string-order correlations that identify the SPT phase, which are invisible to any local classical probe of the raw state.

Classical part: robust decision making

Measurement outcomes are noisy bitstrings. A small classical network trained on labeled examples learns to map noisy statistics to phase labels, absorbing shot noise and device imperfections that would confuse a fixed threshold rule.

Sample efficiency

Because the quantum layers concentrate the signal into a few measured qubits, the classical head needs far fewer shots than tomography-style approaches. In the demo above you can see reliable predictions with only a few hundred shots deep inside a phase.

Enjoy this tool? Build your own with Super