Shivanshu Siyanwal

NMR entanglement

Entanglement Detection via Neural Networks (NMR)

An ANN that detects genuine multipartite entanglement and assigns SLOCC classes on a three-qubit NMR processor, with the canonical form reducing 128 real features to 18.

Published · Quantum Information Processing 2025 QML / NMR / Entanglement

Interactive abstract

Play with the idea

This interactive diagram needs JavaScript. The schematic and write-up below explain the same idea.

Layman view

In plain words

Draft — to be verified. Written in plain language from the paper and project notes on this site.

Entanglement is the quantum link that lets particles share correlations no classical system can. Three qubits can be entangled in different ways, and telling these kinds apart in a real experiment usually means reconstructing the full quantum state, which is costly.

This study trained a neural network to do the sorting. Given the numbers that describe a three-qubit state, it says whether the state is genuinely entangled across all three qubits, and which of six entanglement classes it belongs to.

A key step was shrinking the input. A full description has 128 numbers; writing states in a standard (canonical) form leaves 18, and the trained network needs only a handful of those: 4 for the entanglement test and 6 for the classes. Trained on simulated states, it was tested on 30 states prepared on a real NMR quantum processor.

Expert view

The work, step by step

  1. 01

    Problem

    Detecting genuine multipartite entanglement in experimental NMR systems is hard: witnesses are state-specific and full state tomography is experimentally costly. The trained model needs only a few elements (4 for GME, 6 for SLOCC), which points to reduced-measurement protocols; here they were read off reconstructed states.

  2. 02

    Approach

    An ANN detects genuine multipartite entanglement and assigns one of six SLOCC classes. Each state gives 128 real features (Re and Im of the 64 elements of ρ); writing states in canonical form leaves 18 (14 Re, 4 Im), which are then ranked by ANOVA. The model is trained on simulated states (12 × 10³ states) and tested on 30 experimental three-qubit states prepared on an NMR quantum processor.

  3. 03

    Results

    Validated on 30 experimental three-qubit pseudopure states from an NMR processor: GME detection reached 0.954 ± 0.046 accuracy with 4 features, and SLOCC classification 0.867 ± 0.090 with 6 (overall accuracy over the experimental states, Tables IV–V). The ANN is benchmarked against SVM and KNN, and against existing three-qubit classification schemes (the 3-tangle and correlation tensors).

  4. 04

    Status

    Published in Quantum Information Processing 24(3), 74 (2025); preprint arXiv:2409.19739.

Schematic

How it works

GHZWBS1BS2BS3SEPρ: 64 elements128 real features(Re + Im)canonical form18 (14 Re, 4 Im)ANOVA rankingANNclass label
Schematic. The SLOCC classes of three-qubit pure states that the ANN distinguishes, and the reduction from 128 real features to 18 canonical-form inputs, ranked by ANOVA.

Contribution

My role

I worked on the machine-learning side of this study. I explored multiple ANN architectures and configurations, problem formulations, input representations and encodings (one-hot, binary and categorical), and came up with the feature-reduction approach: writing states in canonical form, then ranking the remaining density-matrix elements by ANOVA. V. Gulati handled the experimental side: sample preparation, NMR gate preparation, the physical experiments and measurements, and the analysis and cleaning of the raw experimental data.

  1. Sample and NMR gate preparationV. Gulati
  2. NMR experiments and measurementV. Gulati
  3. Data analysis and cleaningV. Gulati
  4. Canonical-form features + ANOVAMe (my part)
  5. ANN design and encodingsMe (my part)
My partCollaborators

Publication

The paper

2025 · Journal article

ANN-enhanced detection of multipartite entanglement in a three-qubit NMR quantum processor

V. Gulati*, S. Siyanwal*, Arvind, K. Dorai

Quantum Information Processing, 24(3), 74, 2025

doi:10.1007/s11128-025-04696-8arXiv:2409.19739

Abstract

An artificial neural network identifies which of the six SLOCC classes an experimentally generated three-qubit pseudopure state belongs to, and detects genuine multipartite entanglement. The 128 real features (Re and Im of the 64 elements of ρ) are reduced to 18 (14 Re, 4 Im) by the canonical form, then ranked by ANOVA; the model is trained on simulated states, validated on 30 experimental states from an NMR quantum processor, and benchmarked against SVM and KNN.

* Equal contribution.

More threads