Representability-Aware Neural Networks for Reduced Density Matrices: Application to Fractional Chern Insulators
A new neural network framework has been developed to predict two-particle reduced density matrices (2-RDMs) with a focus on representability conditions. This framework has been applied to study fractional Chern insulators, specifically in the context of twisted bilayer MoTe$_2$. The results indicate that the neural network achieves high accuracy while using significantly fewer parameters compared to traditional methods.
- ▪The neural network framework incorporates representability conditions through its architecture and loss function.
- ▪It can predict 2-RDMs on large momentum meshes by interpolating results from smaller meshes.
- ▪The best-performing neural network achieved an accuracy of 97.07%-98.18% relative to exact diagonalization results.
Opening excerpt (first ~120 words) tap to expand
Condensed Matter > Strongly Correlated Electrons arXiv:2605.20326 (cond-mat) [Submitted on 19 May 2026] Title:Representability-Aware Neural Networks for Reduced Density Matrices: Application to Fractional Chern Insulators Authors:Justin B. Hart, Awwab A. Azam, Thomas Li, Yunxuan Li, Ye Bi, Haining Pan, Jiabin Yu View a PDF of the paper titled Representability-Aware Neural Networks for Reduced Density Matrices: Application to Fractional Chern Insulators, by Justin B. Hart and 6 other authors View PDF HTML (experimental) Abstract:We develop a representability-aware and interpolable neural network (NN) framework for predicting two-particle reduced density matrices (2-RDMs).
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
