An open API service indexing awesome lists of open source software.

https://github.com/ternaus/machine_learning_sign

Machine learning as a way to overcome sign problem.
https://github.com/ternaus/machine_learning_sign

Last synced: about 2 months ago
JSON representation

Machine learning as a way to overcome sign problem.

Awesome Lists containing this project

README

          

# Extrapolation to the region with a sign problem.

## Introduction

Sign problem is a fundamential problem in a theoretical condensed matter physics that pervents theorist from numerical investigation properties of the fermionic Hubbard model. There is no clear way how to measure quantities of interest in a region where average sign in the monte carlo simulations is close to zero, but I will try to use machine learning algorithms to create a model that will be able to extrapolate to the regions of interest.

Sign problem will not appear at:
* mu = 0 This corresponds to a half filled case, in which sign = 1 protected by the particle symmatric form of the Hamiltonian.
* U = 0 In this case there is no sign problem and solution to the problem can be found analytically.

In all other cases:

Sign will appear at:
* U -> infinity
* beta -> infinity

## What is my training data?

As a training set I will use output of the Determinant Monte Carlo simulations generated by QUEST package and exact diagonalization results generated by the ALPS package.

### What lattice I am working with?

In this project I will work only with the square lattice, although other lattices like honeycomb,triangular, Kagome, Lieb and others may be investigated later.

I will imply periodic boundary conditions on the lattice and I will not use lattices with sizes less than 4 in any direction.

### What are my input parameters?

As an input parameters I will use:

* Nx - Number of sites in the x direction. (even numbers only to avoid frustration)
* Ny - Number of sites in the y direction. (even numbers only to avoid frustration)
* mu - chemical potential
* beta - inverse temperature
* U - interaction strength

### What data do we have?

Exact diagonalization will give you results for any mu and U, but for (Nx, Ny) = (4, 4) and T = 0 (beta = infinity)

Quantum Monte Carlo data will be give for different U, beta, mu, Nx, Ny. Data files where < 0.1 will be considered bad (not reliable) and they will be exluded from the consideration.

### What algorithms will I use?

I am planning on using all regression algorithms from the scikit-learn package.
Comparison of their efficiency will be also provided.