Neural Network Predicts Quantum States for Fractional Chern Insulators
Researchers have developed a representability-aware neural network framework to predict two-particle reduced density matrices (2-RDMs) for fractional Chern insulators. The network incorporates representability conditions into its architecture and loss function, and can interpolate across momentum meshes. Applied to twisted bilayer MoTe₂ at a twist angle of 3.89° and hole filling 2/3, the model was trained on exact-diagonalization 2-RDMs from meshes with 12 or 18 momentum points using six different architectures. The framework can predict 2-RDMs on larger meshes or serve as a variational ansatz optimized by energy minimization.
Key facts
- Neural network predicts 2-RDMs for fractional Chern insulators
- Network incorporates representability conditions in architecture and loss function
- Operates on different momentum meshes with interpolated representability condition
- Applied to twisted bilayer MoTe₂ at twist angle 3.89° and hole filling 2/3
- Trained on exact-diagonalization 2-RDMs from meshes with 12 or 18 momentum points
- Six different NN architectures were tested
- Framework can interpolate exact results from small to large meshes
- Can be used as variational 2-RDM ansatz optimized by energy minimization
Entities
Institutions
- arXiv