This fitting function models the inelastic part of the dynamic structure factor for a particle undergoing continuous and isotropic rotational diffusion [1], IsoRotDiff.
where:
Because of the spherical symmetry of the problem, the structure factor is expressed in terms of the spherical Bessel functions.
Name | Type | Default | Description |
---|---|---|---|
N | |||
Q |
(double, default=0.3) Momentum transfer
(integer, default=25) The default N=25 assures normalization condition
with three significant digits for , a comfortable upper bound for the vast majority of QENS data.
Example - Global fit to a synthetic inelastic signal:
The signal is modeled by the convolution of a resolution function with the inelastic component of a rotator. The resolution is modeled as a normal distribution. We insert a random noise in the rotator. Finally, we choose a linear background noise. The goal is to find out the radius of the rotator. the relaxation time, and the overal intensity of the signal with a fit to the following model:
import numpy as np
try:
from scipy.special import spherical_jn
def sjn(n, z): return spherical_jn(range(n+1), z)
except ImportError:
from scipy.special import sph_jn
def sjn(n, z): return sph_jn(n, z)[0]
"""Generate resolution function with the following properties:
1. Gaussian in Energy
2. Dynamic range = [-0.1, 0.1] meV with spacing 0.0004 meV
3. FWHM = 0.005 meV
"""
dE=0.0004; FWHM=0.005; sigma = FWHM/(2*np.sqrt(2*np.log(2)))
dataX = np.arange(-0.1,0.1,dE); Emin=min(dataX); Emax=max(dataX); nE=len(dataX)
rdataY = np.exp(-0.5*(dataX/sigma)**2) # the resolution function
Qs = np.array([0.3, 0.5, 0.7, 0.9, 1.1, 1.3, 1.5, 1.9]) # Q-values
nQ = len(Qs)
resolution=CreateWorkspace(np.tile(dataX,nQ), np.tile(rdataY,nQ), NSpec=nQ, UnitX="deltaE",
VerticalAxisUnit="MomentumTransfer", VerticalAxisValues=Qs)
"""Generate a synthetic inelastic signal for a particle undergoing isotropic rotational diffusion.
1. Radius of rotation = 2.5 Angstroms
2. Relaxation time = 187 ps
3. Up to 10% of noise in the quasi-elastic signal
4. Linear background noise, up to 1% of the quasi-elastic intensity
"""
R=2.5; tau=187.0; hbar=0.658211626 # hbar units are ps*meV
N=25 # number of harmonics in the inelastic signal
qdataY=np.empty(0) # will hold all Q-values (all spectra)
H=2-np.random.random() # global intensity
for Q in Qs:
centre=dE*np.random.random() # some shift along the energy axis
dataY=np.zeros(nE) # holds the inelastic signal for this Q-value
js=sjn(N,Q*R) # spherical bessel functions from L=0 to L=N
for L in range(1,N+1):
HWHM = L*(L+1)*hbar/tau; aL=(2*L+1)*js[L]**2
dataY += H*aL/np.pi * HWHM/(HWHM**2+(dataX-centre)**2) # add component
dataY = dE*np.convolve(rdataY, dataY, mode="same") # convolve with resolution
noise = dataY*np.random.random(nE)*0.1 # noise is up to 10% of the inelastic signal
background = np.random.random()+np.random.random()*dataX # linear background
background = (0.01*H*max(dataY)) * (background/max(np.abs(background))) # up to 1%
qdataY=np.append(qdataY, dataY+background)
data=CreateWorkspace(np.tile(dataX,nQ), qdataY, NSpec=nQ, UnitX="deltaE",
VerticalAxisUnit="MomentumTransfer", VerticalAxisValues=Qs)
"""Our fitting model is:
S(Q,E) = Convolution(resolution, InelasticIsoRotDiff) + LinearBackground
We do a global fit (all spectra) to the synthetic data workspace to find out
the global intensity H, the radius R, and the relaxation time tau.
"""
# This is the template fitting model for each spectrum (each Q-value):
single_model_template="""(composite=Convolution,FixResolution=true,NumDeriv=true;
name=TabulatedFunction,Workspace=resolution,WorkspaceIndex=_WI_,Scaling=1,Shift=0,XScaling=1;
name=InelasticIsoRotDiff,N=25,Q=_Q_,Height=1,Radius=0.98,Tau=10,Centre=0,
constraints=(0<Height,0.1<Radius,0.1<Tau));
name=LinearBackground,A0=0,A1=0"""
# Now create the string representation of the global model (all spectra, all Q-values):
global_model="composite=MultiDomainFunction,NumDeriv=true;"
wi=0
for Q in Qs:
single_model = single_model_template.replace("_Q_", str(Q)) # insert Q-value
single_model = single_model.replace("_WI_", str(wi)) # workspace index
global_model += "(composite=CompositeFunction,NumDeriv=true,$domains=i;{0});\n".format(single_model)
wi+=1
# The Height, Radius, and Tau are the same for all spectra, thus tie them:
ties=['='.join(["f{0}.f0.f1.Radius".format(wi) for wi in reversed(range(nQ))]),
'='.join(["f{0}.f0.f1.Height".format(wi) for wi in reversed(range(nQ))]),
'='.join(["f{0}.f0.f1.Tau".format(wi) for wi in reversed(range(nQ))]) ]
global_model += "ties=("+','.join(ties)+')' # insert ties in the global model string
# Now relate each domain(i.e. spectrum) to each single model
domain_model=dict()
for wi in range(nQ):
if wi == 0:
domain_model.update({"InputWorkspace": data.name(), "WorkspaceIndex": str(wi),
"StartX": str(Emin), "EndX": str(Emax)})
else:
domain_model.update({"InputWorkspace_"+str(wi): data.name(), "WorkspaceIndex_"+str(wi): str(wi),
"StartX_"+str(wi): str(Emin), "EndX_"+str(wi): str(Emax)})
# Invoke the Fit algorithm using global_model and domain_model:
output_workspace = "glofit_"+data.name()
Fit(Function=global_model, Output=output_workspace, CreateOutput=True, MaxIterations=500, **domain_model)
# Extract Height, Radius, and Tau from workspace glofit_data_Parameters, the output of Fit:
nparms=0
parameter_ws = mtd[output_workspace+"_Parameters"]
for irow in range(parameter_ws.rowCount()):
row = parameter_ws.row(irow)
if row["Name"]=="f0.f0.f1.Radius":
Radius=row["Value"]
nparms+=1
elif row["Name"]=="f0.f0.f1.Height":
Height=row["Value"]
nparms+=1
elif row["Name"]=="f0.f0.f1.Tau":
Tau=row["Value"]
nparms+=1
if nparms==3:
break # We got the three parameters we are interested in
# Check nominal and optimal values are within error ranges:
if abs(H-Height)/H < 0.1:
print("Optimal Height within 10% of nominal value")
if abs(R-Radius)/R < 0.05:
print("Optimal Radius within 5% of nominal value")
if abs(tau-Tau)/tau < 0.1:
print("Optimal Tau within 10% of nominal value")
Output:
Optimal Height within 10% of nominal value
Optimal Radius within 5% of nominal value
Optimal Tau within 10% of nominal value
Categories: FitFunctions | QuasiElastic
C++ source: InelasticIsoRotDiff.cpp (last modified: 2017-01-16)
C++ header: InelasticIsoRotDiff.h (last modified: 2016-10-11)