[Paper Review] Quantum neural network autoencoder and classifier applied to an industrial case study

Stefano Mangini, Alessia Marruzzo+4 more|arXiv (Cornell University)|Jan 1, 2022

Quantum Computing Algorithms and Architecture58 references26 citations

TL;DR

This paper proposes a hybrid quantum-classical pipeline combining a variational quantum autoencoder and a quantum classifier to compress and classify real industrial data from a water/oil separator in an Eni oil treatment plant. Using only 6 trainable parameters, the quantum pipeline achieves 87.4% classification accuracy on simulated data and 82.3% on real IBM quantum hardware—comparable to classical methods—demonstrating practical feasibility of quantum machine learning on near-term devices with real-world industrial data.

ABSTRACT

Quantum computing technologies are in the process of moving from academic research to real industrial applications, with the first hints of quantum advantage demonstrated in recent months. In these early practical uses of quantum computers it is relevant to develop algorithms that are useful for actual industrial processes. In this work we propose a quantum pipeline, comprising a quantum autoencoder followed by a quantum classifier, which are used to first compress and then label classical data coming from a separator, i.e., a machine used in one of Eni's Oil Treatment Plants. This work represents one of the first attempts to integrate quantum computing procedures in a real-case scenario of an industrial pipeline, in particular using actual data coming from physical machines, rather than pedagogical data from benchmark datasets.

Motivation & Objective

• To demonstrate the feasibility of applying quantum machine learning algorithms to real industrial data from operational equipment, rather than synthetic or benchmark datasets.
• To evaluate whether a quantum autoencoder can effectively compress high-dimensional classical data from a real industrial separator into a lower-dimensional quantum latent space.
• To compare the performance of a quantum classifier trained on compressed quantum data against classical baselines in a binary classification task.
• To validate the robustness of the quantum pipeline on actual NISQ-era quantum hardware, including noise and measurement stochasticity.
• To assess the efficiency of the quantum approach in terms of trainable parameters compared to classical models.

Proposed method

• Implement a variational quantum autoencoder (VQA) with a parameterized quantum circuit to compress 4D classical input data (oil level, flow, pressure, valve opening) into a 2D latent quantum state.
• Use amplitude encoding to load classical data into the quantum state via single-qubit rotations, enabling non-linear encoding of input features.
• Train the quantum autoencoder using a hybrid quantum-classical optimization loop, minimizing the reconstruction error between original and decoded quantum states.
• Evaluate performance using quantum fidelity and average reconstruction error (¯e ≈ 5%) to validate compression quality.
• Train a quantum classifier (variational quantum circuit) on the latent quantum state to perform binary classification based on KMeans clustering labels.
• Test the full pipeline on both ideal quantum simulation and real IBM quantum hardware (ibmqx2), using 1024 shots per circuit and averaging over multiple executions.

Experimental results

Research questions

• RQ1 Can a variational quantum autoencoder effectively compress real industrial data from a physical separator into a lower-dimensional quantum latent space with performance comparable to classical autoencoders?
• RQ2 Does a quantum classifier trained on the compressed quantum state achieve classification accuracy on par with classical models when applied to real industrial data?
• RQ3 How robust is the quantum pipeline to noise and measurement stochasticity in near-term quantum hardware?
• RQ4 Can the quantum pipeline achieve comparable performance to classical models using significantly fewer trainable parameters (6 vs. 16)?
• RQ5 Is the quantum machine learning pipeline viable and executable on currently available NISQ devices when applied to real-world industrial data?

Key findings

• The quantum autoencoder achieved an average reconstruction error of approximately 5%, matching the performance of the classical autoencoder.
• Quantum fidelity between original and decoded states remained high, confirming effective compression and recovery of the input data in the quantum regime.
• The quantum classifier achieved 87.4% classification accuracy on a test set of 1,000 samples in simulation, comparable to the classical KNN classifier (89.7%) on normalized data.
• On real IBM quantum hardware (ibmqx2), the quantum classifier achieved 82.3% accuracy with a standard deviation of ±1.3%, demonstrating robustness to hardware noise and measurement errors.
• The quantum pipeline required only 6 trainable parameters, significantly fewer than the 16 used in the classical model, indicating improved parameter efficiency on NISQ devices.
• The results confirm that quantum machine learning pipelines can be successfully deployed on real industrial data using current quantum hardware, marking a foundational step toward scalable industrial quantum applications.