NeuralQuantumStates.jl

Mon, 10 Aug 2026 00:00:00 +0000

NeuralQuantumStates.jl trains neural quantum states with variational Monte Carlo in Julia. It covers the full workflow: define a lattice and its symmetries, build a reduced basis, compute local energies, pick a neural network (RBM, symmetric RBM, or Jastrow, built on Lux.jl), sample with Metropolis, and optimise with stochastic reconfiguration. The closest Python equivalents are NetKet and jVMC.

The code is split into separate packages so you can take only what you need. LatticeSpaceGroups.jl and ConnectedBasisConfigurations.jl handle lattices and local energies with no machine learning involved, so they work on their own. NQSCore.jl, NQSAnsatze.jl, NQSSamplers.jl, and NQSOptimisers.jl add the neural network layer on top. Symmetries come from , operators from OperatorAlgebra.jl.

Useful for: running variational Monte Carlo on closed quantum many-body systems in Julia, or replacing one piece of that pipeline without rewriting the rest.

Demonstrates: splitting a large Julia project into separate packages, automatic differentiation, Monte Carlo sampling, natural-gradient optimisation, and checking results against a method that gives the exact answer.

Why it matters: training a quantum state by sampling it and following the gradient is the same loop that variational algorithms run on real quantum hardware.

Status: the pipeline works, and its results match exact diagonalization. Still to come: GPU support (CUDA, AMDGPU, Metal), running across multiple machines with MPI, and time evolution. The API will still change.

Machine Learning |

NeuralQuantumStates.jl