AI for Science and AI for mathematics are now major international research projects. One aspect of AI for science is modelling systems using machine learning which in the past were modelled using classical tools like ODE or PDE. Classical experimental science typically incorporates both an explanatory model and a probabilistic/statistical model for estimating model parameters with experiments. We make a foray into the latter problem by developing a machine learning system in pytorch for sampling the Gibbs distribution of the Ising model. We use the classical metropolis algorithm to generate approximate samples for the Ising model on a 10,000x10,000 grid. Then we train a neural network to estimate the function that maps a configuration of the grid to the results of applying the metropolis algorithm for 1000 steps. The result is a pseudorandom sampler, analogous to the Mersenne Twister. Most pseudorandom samplers approximate the uniform probability distribution. However we are approximating a non-uniform probability distribution. Like a uniform pseudorandom sampler, our neural network produces a cyclical deterministic sequence of elements with a long cycle length. However the frequency distribution for configurations of a cycle will approximate the Gibbs distribution for the Ising model Hamiltonian.