我正在尝试模拟Barabasi-Albert网络中的Ising相变,并尝试复制一些可观测量的结果,如磁化和能量,就像在Ising网格模拟中观察到的那样。但是,我在解释结果时遇到了麻烦:不确定物理是否错误或实施中是否存在错误。以下是最低工作示例:
import numpy as np
import networkx as nx
import random
import math
## sim params
# coupling constant
J = 1.0 # ferromagnetic
# temperature range, in units of J/kT
t0 = 1.0
tn = 10.0
nt = 10.
T = np.linspace(t0, tn, nt)
# mc steps
steps = 1000
# generate BA network, 200 nodes with preferential attachment to 3rd node
G = nx.barabasi_albert_graph(200, 3)
# convert csr matrix to adjacency matrix, a_{ij}
adj_matrix = nx.adjacency_matrix(G)
top = adj_matrix.todense()
N = len(top)
# initialize spins in the network, ferromagnetic
def init(N):
return np.ones(N)
# calculate net magnetization
def netmag(state):
return np.sum(state)
# calculate net energy, E = \sum J *a_{ij} *s_i *s_j
def netenergy(N, state):
en = 0.
for i in range(N):
for j in range(N):
en += (-J)* top[i,j]*state[i]*state[j]
return en
# random sampling, metropolis local update
def montecarlo(state, N, beta, top):
# initialize difference in energy between E_{old} and E_{new}
delE = []
# pick a random source node
rsnode = np.random.randint(0,N)
# get the spin of this node
s2 = state[rsnode]
# calculate energy by summing up its interaction and append to delE
for tnode in range(N):
s1 = state[tnode]
delE.append(J * top[tnode, rsnode] *state[tnode]* state[rsnode])
# calculate probability of a flip
prob = math.exp(-np.sum(delE)*beta)
# if this probability is greater than rand[0,1] drawn from an uniform distribution, accept it
# else retain current state
if prob> random.random():
s2 *= -1
state[rsnode] = s2
return state
def simulate(N, top):
# initialize arrays for observables
magnetization = []
energy = []
specificheat = []
susceptibility = []
for count, t in enumerate(T):
# some temporary variables
e0 = m0 = e1 = m1 = 0.
print 't=', t
# initialize spin vector
state = init(N)
for i in range(steps):
montecarlo(state, N, 1/t, top)
mag = netmag(state)
ene = netenergy(N, state)
e0 = e0 + ene
m0 = m0 + mag
e1 = e0 + ene * ene
m1 = m0 + mag * mag
# calculate thermodynamic variables and append to initialized arrays
energy.append(e0/( steps * N))
magnetization.append( m0 / ( steps * N))
specificheat.append( e1/steps - e0*e0/(steps*steps) /(N* t * t))
susceptibility.append( m1/steps - m0*m0/(steps*steps) /(N* t *t))
print energy, magnetization, specificheat, susceptibility
plt.figure(1)
plt.plot(T, np.abs(magnetization), '-ko' )
plt.xlabel('Temperature (kT)')
plt.ylabel('Average Magnetization per spin')
plt.figure(2)
plt.plot(T, energy, '-ko' )
plt.xlabel('Temperature (kT)')
plt.ylabel('Average energy')
plt.figure(3)
plt.plot(T, specificheat, '-ko' )
plt.xlabel('Temperature (kT)')
plt.ylabel('Specific Heat')
plt.figure(4)
plt.plot(T, susceptibility, '-ko' )
plt.xlabel('Temperature (kT)')
plt.ylabel('Susceptibility')
simulate(N, top)
观察到的结果:
我已经尝试过大量注释代码,如果有什么我忽略了请问。
问题:
答案 0 :(得分:3)
首先,由于这是一个编程站点,让我们分析一下该程序。您的计算效率非常低,这使得探索更大的图形变得不切实际。在您的情况下,邻接矩阵是200x200(40000)元素,只有大约3%的非零。将其转换为密集矩阵意味着更多的计算,同时评估montecarlo
例程中的能量差异和netenergy
中的净能量。以下代码在我的系统上执行速度提高了5倍,预计更大的速度可以使用更大的图表:
# keep the topology as a sparse matrix
top = nx.adjacency_matrix(G)
def netenergy(N, state):
en = 0.
for i in range(N):
ss = np.sum(state[top[i].nonzero()[1]])
en += state[i] * ss
return -0.5 * J * en
注意因子中的0.5 - 因为邻接矩阵是对称的,每对自旋都被计算两次!
def montecarlo(state, N, beta, top):
# pick a random source node
rsnode = np.random.randint(0, N)
# get the spin of this node
s = state[rsnode]
# sum of all neighbouring spins
ss = np.sum(state[top[rsnode].nonzero()[1]])
# transition energy
delE = 2.0 * J * ss * s
# calculate transition probability
prob = math.exp(-delE * beta)
# conditionally accept the transition
if prob > random.random():
s = -s
state[rsnode] = s
return state
请注意转换能量中的因子2.0 - 代码中缺少它!
这里有一些numpy
索引魔法。 top[i]
是节点 i 的稀疏邻接行向量,top[i].nonzero()[1]
是非零元素的列索引(top[i].nonzero()[0]
是行索引,其中都是等于0,因为它是行向量)。因此state[top[i].nonzero()[1]]
是节点 i 的相邻节点的值。
现在去物理学。热力学性质是错误的,因为:
e1 = e0 + ene * ene
m1 = m0 + mag * mag
应该是:
e1 = e1 + ene * ene
m1 = m1 + mag * mag
和
specificheat.append( e1/steps - e0*e0/(steps*steps) /(N* t * t))
susceptibility.append( m1/steps - m0*m0/(steps*steps) /(N* t *t))
应该是:
specificheat.append((e1/steps/N - e0*e0/(steps*steps*N*N)) / (t * t))
susceptibility.append((m1/steps/N - m0*m0/(steps*steps*N*N)) / t)
(你最好在早期平均能量和磁化力)
这使得热容量和对磁场的磁化率为正值。请注意敏感度分母中的单个t
。
现在该程序(希望)正确,让我们谈谈物理。对于每个温度,你从一个全向旋转状态开始,然后让它一次进化一个旋转。显然,除非温度为零,否则该初始状态远离热平衡,因此系统将开始向对应于给定温度的状态空间部分漂移。这个过程被称为热化,并且在此期间收集静态信息是没有意义的。您必须始终将模拟在给定温度下分成两部分 - 热化和实际显着运行。需要多少次迭代才能达到平衡?很难说 - 当它变得(相对)稳定时,使用能量的移动平均值并进行监控。
其次,更新算法每次迭代改变一次旋转,这意味着程序将非常缓慢地探索状态空间,并且您需要大量的迭代才能获得分区函数的良好近似。通过200次旋转,1000次迭代可能就够了。
其余的问题实际上并不属于这里。