Bose-Hubbard model

NetKet can be applied also to bosonic systems. To get acquainted with this kind of calculations, you can have a look at the tutorial on the one-dimensional Bose-Hubbard model, in Tutorials/BoseHubbard1d.

There, we study the Hamiltonian

where the hopping term runs over pairs of nearest-neighbors on the one-dimensional lattice.

The steps to be followed are very similar to the spin Hamiltonians we have seen so far, but for completeness we repeat them here.

Input file

The Python script can be used to set up the JSON input file for the NetKet executable. In the following we go through this script step by step, explaining the several fields.

Defining the lattice

In this section of the input we specify the graph on which our spins live.

    'Name'           : 'Hypercube',
    'L'              : 12,
    'Dimension'      : 1 ,
    'Pbc'            : True,

The name of the parameters should be self-explanatory, for example here we are taking a Hypercube in one Dimension with periodic boundaries (Pbc) and with a linear extent of L 12 sites. If you wanted to study instead a square lattice, you would just set Dimension : 2.

Apart from Built-in Graphs (such as the Hypercube), you can easily specify virtually any other custom graph, as explained here.

Defining the Hamiltonian

Next, we specify the Hamiltonian we want to simulate.

    'Name'           : 'BoseHubbard',
    'U'              : 4.0,
    'Nmax'           : 3,
    'Nbosons'        : 12,

Here we specify the name of the Hamiltonian and also the fact the we want to study, picking one of the Built-in Hamiltonians. Notice that here we specify the maximum number of bosons allowed per site (with the field Nmax) and also specify that we want the ground-state in the sector with total number of particles fixed, in this specific case we work at unit filling, i.e. .

Further notice that NetKet allows to define custom Hamiltonians, simply working at the level of input files, as explained here.

Defining the Machine

In this section of the input we specify what wave function ansatz we wish to use. Here, we take a Restricted Boltzmann Machine RbmSpinSymm with spin hidden units and permutation symmetry. Since we are working with a translation-invariant Hamiltonian, and we are interested in the , zero momentum ground-state, this is a sensible choice. To further use this machine we must also specify the number of hidden units we want to have. In this machine we also must set Alpha, where , as done in the example input.

    'Name'           : 'RbmSpinSymm',
    'Alpha'          : 4.0,

Further details about the Restricted Boltzmann Machines and the other machines implemented in NetKet can be found here.

Defining the Sampling scheme

Another crucial ingredient for the learning part is the Markov-Chain Monte Carlo scheme used for sampling. Here, we consider a Metropolis sampler implementing Hamiltonian moves (see here for a description of this specific family of sampler).

    'Name'           : 'MetropolisHamiltonian',

An important reason to chose this sampler in this case is that we want to make sure to preserve all the symmetries of the Hamiltonian during the sampling. Basically, what the sampler does in this case is that it choses a pair of neighboring spins at random and proposes an hopping of particles, if allowed.

This is crucial for example if we want our specification 'Nbosons' : 12 to be verified at all times during the sampling. If instead of Hamiltonian moves we chose local Metropolis moves, during the sampling our total number of particles would fluctuate, thus violating the wanted constraint. In that case however you can specify a chemical potential to control the average number of particles.

Defining the Learning scheme

Finally, we must specify what learning algorithm we wish to use. Together with the choice of the machine, this is the most important part of the simulation. The method of choice in NetKet is the Stochastic Reconfiguration Sr, developed by S. Sorella and coworkers. For an introduction to this method, you can have a look at the book (2). The code snippet defining the learning methods is:

    'Method'         : 'Sr',
    'Nsamples'       : 1.0e3,
    'NiterOpt'       : 4000,
    'Diagshift'      : 5.0e-3,
    'UseIterative'   : False,
    'OutputFile'     : 'test',

Also, notice that we need to specify an optimizer. Here we choose AdaMax with default parameters, specifying the following section of the input:

    'Name'           : 'AdaMax',

More details about the optimizers can be found here, whereas learning algorithms to find the ground state are discussed here.

Running the simulation

Once you have finished preparing the input file in python, you can just run:


this will generate a JSON file called bosehubbard1d.json ready to be fed to the NetKet executable. At this point then you can just run

netket bosehubbard1d.json

if you want to run your simulation on a single core, or

mpirun -n NP netket bosehubbard1d.json

if you want to run your simulation on NP cores (changes NP to the number of cores you want to use).

At this point, the simulation will be running and log files will be generated in real time, until NetKet finishes its tasks.

Output files

Since in the Learning section we have specified 'OutputFile' : "test", two output files will be generated with the “test” prefix, i.e. test.log, a JSON file containing the results of the learning procedure as it advances, and containing backups of the optimized wave function.

For each iteration of the learning, the output log contains important information which can visually inspected just opening the file.


For example, you can see here that we have the expectation value of the energy (Mean), its statistical error (Sigma), and an estimate of the autocorrelation time (Taucorr). Apart from the Energy, the learning algorithm also records the EnergyVariance, namely which is smaller and smaller when converging to an exact eigenstate of the Hamiltonian.

If you want, you can also plot these results while the learning is running, just using the convenience script:


An example result is shown below, where you can see that the energy would converge to the exact result during the learning.

Responsive image

It is also an interesting exercise to compare the accuracy on the energies obtained here with the Restricted Boltzmann machine with the accuracy obtained with Feed-Forward neural networks on the same system, for example in Reference (3).


  1. Carleo, G., & Troyer, M. (2017). Solving the quantum many-body problem with artificial neural networks. Science, 355 602
  2. Becca, F., & Sorella, S. (2017). Quantum Monte Carlo Approaches for Correlated Systems. Cambridge University Press.
  3. Saito, H., & Kato, M. (2018). Machine learning technique to find quantum many-body ground states of bosons on a lattice. J. Phys. Soc. Jpn. 87, 014001