Equivariant Continuous Normalizing Flows with Offline Sampling for Fermionic Ground State Estimation
The framework uses a continuous normalizing flow (CNF) to refine a fixed antisymmetric base wavefunction for fermionic variational Monte Carlo (VMC). By using permutation-equivariant neural ODEs and offline sampling, exact samples from the flow's Born distribution are obtained without MCMC during training. This allows for efficient, parallel training on multiple GPUs, achieving ground state energies competitive with high-level methods.