Entanglement and discord classification via deep learning

In this work, we propose a deep learning-based approach for quantum entanglement and discord classification using convolutional autoencoders. We train models to distinguish entangled from separable bipartite states for $d \times d$ systems with local dimension $d$ ranging from two to seven, which enables identification of bound and free entanglement. Through extensive numerical simulations across various quantum state families, we demonstrate that our model achieves high classification accuracy. Furthermore, we leverage the learned representations to generate samples of bound entangled states, the rarest form of entanglement and notoriously difficult to construct analytically. We separately train the same convolutional autoencoders architecture for detecting the presence of quantum discord and show that the model also exhibits high accuracy while requiring significantly less training time.

💡 Research Summary

The paper introduces a deep‑learning framework based on convolutional autoencoders (CAEs) for the classification of quantum entanglement and quantum discord in bipartite systems of dimension (d\times d) with (d) ranging from 2 to 7. For entanglement detection, the authors adopt an unsupervised approach: they generate a large training set consisting solely of separable states, constructed as random convex combinations of tensor products of local mixed states. The CAE is trained to reconstruct these density matrices, minimizing a Frobenius‑norm reconstruction loss. After training, the reconstruction error serves as a direct indicator of non‑separability: states that yield a high error are flagged as entangled. This simple criterion cleanly separates negative‑partial‑transpose (NPT) states from separable ones, achieving >98 % accuracy for all dimensions tested.

A notable challenge is the detection of bound entangled states—states that are PPT yet non‑separable. Initially, the model misclassifies such states because they are often represented as sparse matrices, whereas the training data are dense. The authors resolve this by applying random local unitary transformations to convert sparse representations into dense ones before feeding them to the CAE. This preprocessing restores high classification performance on bound entangled samples across all dimensions.

Beyond classification, the authors exploit the learned latent space of the CAE to generate new bound entangled states. By sampling latent vectors and decoding them back to density matrices, they obtain candidate states that are subsequently verified using the symmetric‑extension criterion. This demonstrates that the autoencoder not only discriminates but also serves as a generative model for rare quantum resources.

For quantum discord detection, the same CAE architecture is employed in a supervised setting. The training set comprises only classical‑classical (CC) states, which have zero discord. The model learns to distinguish CC states from states that possess discord (classical‑quantum or quantum‑classical). Remarkably, the discord classifier converges after a single epoch and reaches >99 % accuracy for (d\ge 3), while requiring significantly less training time than the entanglement model.

The paper situates its contributions within prior work on machine‑learning‑based entanglement detection, highlighting that earlier approaches often used more complex architectures (e.g., pseudo‑siamese networks, GANs) or relied on measurement data rather than raw density matrices. In contrast, the presented method is streamlined, operates directly on the density matrix, and scales efficiently to higher dimensions. The authors also discuss limitations: the need for full state tomography to obtain density matrices in experimental settings, and the untested scalability beyond (d=7).

In summary, the study offers a unified, efficient deep‑learning pipeline that (i) classifies entangled versus separable states via reconstruction error, (ii) correctly identifies bound entanglement through preprocessing and latent‑space manipulation, (iii) generates novel bound‑entangled states, and (iv) detects quantum discord with minimal training effort. The results suggest that convolutional autoencoders can capture fundamental structural features of quantum states, providing a promising tool for quantum resource identification and synthesis. Future work will explore higher dimensions, noisy experimental data, and extensions to other quantum resources such as coherence and magic.