Enhanced sampling in generalized ensemble with large gap of sampling parameter: case study in temperature space random walk
We present an efficient sampling method for computing a partition function and accelerating configuration sampling. The method performs a random walk in the $λ$ space, with $λ$ being any thermodynamic variable that characterizes a canonical ensemble such as the reciprocal temperature $β$ or any variable that the Hamiltonian explicitly depends on. The partition function is determined by minimizing the difference of the thermal conjugates of $λ$ (the energy in the case of $λ=β$), defined as the difference between the value from the dynamically updated derivatives of the partition function and the value directly measured from simulation. Higher-order derivatives of the partition function are included to enhance the Brownian motion in the $λ$ space. The method is much less sensitive to the system size, and the size of $λ$ window than other methods. On the two dimensional Ising model, it is shown that the method asymptotically converges the partition function, and the error of the logarithm of the partition function is much smaller than the algorithm using the Wang-Landau recursive scheme. The method is also applied to off-lattice model proteins, the $AB$ models, in which cases many low energy states are found in different models.