426 KiB
Showcasing of AMBER utilizing synthetic neutron scattering data¶
import numpy as np from AMBER.background import background import matplotlib.pyplot as plt
Generate signal data¶
The data for this tutorial is generated using the expected neutron scattering signal in an inelastic exmperiment measuring MnF$_2$. The dispersion relation, i.e. the energy at a give (H,K,L) position is given by the analytical formula Yamany et al. 2010:
$$\omega(H,K,L) = \sqrt{\left(2S z_2 J_2+D+2 z_1 S J_1 \sin{L \pi}^2\right)^2-\left(2 S z_2 J_2 \cos(H \pi)\cos(K \pi) \cos(L \pi)\right)^2}$$
In the above $J_1$, $J_2$, and $D$ are the magnetic coupling strengths and single-ion anisotropy which determine the amplitude of the dispersion. The parameters $S = 5/2$, $z_1 = 2$, and $z_2 = 8$ corresponds the the size of the magnetic spin, and the number of nearest and next-nearest neighbours in the MnF$_2$ lattice.
From the above equation, the imporant fact is that it depends on the 3D vector $Q = (H,K,L)$. For simplicity, we assume that $K$ is zero and thus reduce the data to 3D, i.e. $(H,L,\omega)$
Due to how neutron scattering experiments function, the above dispersion will not be infinitely thin but rather extended. Here, this is replicated by a simple gaussian smearing along $\omega$
# Define parameters, dispersion, and Smearing S = 5/2 z1 = 2 z2 = 8 def SpinWave(Q,J1,J2,D): return np.sqrt(np.power(2*S*z2*J2+D+2*z1*S*J1*np.sin(Q[2]*np.pi)**2,2.0)- np.power(2*S*z2*J2*np.cos(Q[0]*np.pi)*np.cos(Q[1]*np.pi)*np.cos(Q[2]*np.pi),2.0)) def Intensity(H,K,L,E): sigmaE = 0.25 omega = SpinWave(np.array([H,K,L]), J1=0.0354,J2=0.1499,D=0.131 ) I = np.exp(-np.power(omega-E,2.0)/(2*sigmaE**2)) return I
Define 3D grid domain¶
The input to AMBER is a 3D cube of intensities which is define below
# Data will ge simulated along H and L but with K = 0 h = np.linspace(-0.1,2.1,101) k = 0 l = np.linspace(-0.1,2.1,101) # Choose a sufficient energy range e = np.linspace(0.5,8,61) # Generate the grid upon which the dispersion is calculated H,L,E = np.meshgrid(h,l,e) K = np.zeros_like(H) # Intensities are calculated with the smearing and scaled I = 30*Intensity(H,K,L,E) # Poisson noice is added to the intensity I = np.random.poisson(I).astype(float)
Introduction of NaN-values¶
Triple axis instruments cannot measure all lengths of $Q$. this we can mimic by exchaning the intensity at these points by NaNs
QLength = np.linalg.norm([H,L],axis=0) I[QLength<0.35] = np.nan I[QLength>2.5] = np.nan
Add background¶
The main feature of AMBER is the background segmentation, which requires the data to have a background. As describe in the article, background is assumed to:
- Rotation independence of the background.
- Smooth change of background along energy and $|\vec{Q}|$.
- The signal is sparse but continuous in energy and $|\vec{Q}|$.
In the following, a background is generated with a higher amplitude for low and high $|\vec{Q}|$ mimicing instrumental background artefacts from triple axis neutron experiments - these corresponds to the direct beam contribution and to increased background at larger scattering angles.
# Background definition def background_simulation(q,amplitude,gamma,mu,amplitude2,gamma2): return amplitude*((gamma/(q**2+gamma**2))+amplitude2*np.exp(-np.power(q-mu,2.0)/(2*gamma2**2))) Q = np.linalg.norm([H,K,L],axis=0) # Choose suitable valies gamma = 0.5 mu = 3.0 amplitude = 20 amplitude2 = 1.0 gamma2 = 0.5 # Generate an example of the background for visual inspection q = np.linspace(0.25,Q.max(),201) bg_test = background_simulation(q,amplitude,gamma,mu,amplitude2,gamma2) # display background amplitude fig,ax = plt.subplots() ax.plot(q,bg_test) ax.set_xlabel('|Q|') ax.set_ylabel('Background intensity') ## Add background to data bg_tmp = background_simulation(Q,amplitude,gamma,mu,amplitude2,gamma2) I+=np.random.poisson(bg_tmp)
Plot a constant energy plot¶
The now background affected data is plotted to show a "before" picture. This is done by choosing a specific energy ($\omega \sim 4.5$ meV) and plotting the intensity as a color map
fig,ax = plt.subplots() # Find the closest energy slice energy = 4.5 EIdx = np.argmin(np.abs(E-energy)) ax.pcolormesh(H[:, :, EIdx], L[:, :, EIdx], I[:, :, EIdx], vmin=0, vmax = 50) ax.set_title('E = {:.2f} meV'.format(e[EIdx])) ax.set_xlabel('H [r.l.u.]') ax.set_ylabel('L [r.l.u.]')
Text(0, 0.5, 'L [r.l.u.]')
Plot a constant H map¶
fig,ax = plt.subplots() h_value = 0.25 HIdx = np.argmin(np.abs(h-h_value)) ax.pcolormesh(L[:, HIdx, :],E[:, HIdx, :],I[:, HIdx, :], vmin=0, vmax=50) ax.set_xlabel('L [r.l.u.]') ax.set_ylabel('E [meV]')
Text(0, 0.5, 'E [meV]')
Run denoising algorithm¶
# Initialize the AMBER background object AMBER = background(dtype=np.float32) # Set the grid sizes AMBER.set_gridcell_size(dqx = 0.022, dqy = 0.022, dE = 0.125) # Alternatively these can be set from the h, l, and e arrays like # AMBER.set_volume_from_limits([h[0],l[0],e[0]],[h[-1],l[-1],e[-1]],) # Input the data AMBER.set_binned_data(h, l, e, I) bins = int((q.max()-q.min())/0.022) # define maximum radius and number of bins AMBER.set_radial_bins(q.max(),n_bins=bins)
Set algorithm parameters ($\lambda, \beta, \mu$)¶
lambda and mu will be determined as described in the paper and we select beta using cross validation
beta_ is selected using cross validation, i.e. mask out the top q quantile of intensity. The beta value for the lowest Root-Mean-Square-Error is then chosen
lambda_tmp = AMBER.MAD_lambda() mu_tmp = AMBER.mu_estimator() beta_range_tmp = np.array([0.1,1.0,10.0,100.0,200.0,300.0,400.0,500.0]) rmse = AMBER.cross_validation(q=0.3,beta_range= beta_range_tmp, lambda_=lambda_tmp, mu_=mu_tmp,n_epochs=15,verbose=False) beta_tmp = beta_range_tmp[np.argmin(rmse)]
Test - ( 0.1 ) RMSE - ( 4.4478 0.1 55.50635141061015 ) : 3.4844415 Test - ( 1.0 ) RMSE - ( 4.4478 1.0 55.50635141061015 ) : 3.4835355 Test - ( 10.0 ) RMSE - ( 4.4478 10.0 55.50635141061015 ) : 3.4762373 Test - ( 100.0 ) RMSE - ( 4.4478 100.0 55.50635141061015 ) : 3.4504814 Test - ( 200.0 ) RMSE - ( 4.4478 200.0 55.50635141061015 ) : 3.443771 Test - ( 300.0 ) RMSE - ( 4.4478 300.0 55.50635141061015 ) : 3.4432797 Test - ( 400.0 ) RMSE - ( 4.4478 400.0 55.50635141061015 ) : 3.4455476 Test - ( 500.0 ) RMSE - ( 4.4478 500.0 55.50635141061015 ) : 3.44926
Run the denoising algorithm using the parameters obtained using the heuristic¶
# set number of epochs n_epochs = 20 AMBER.denoising(AMBER.Ygrid,lambda_tmp,beta_tmp,mu_tmp,n_epochs,verbose=True)
Iteration 1 Loss function: 24608190.0 Iteration 2 Loss function: 18320822.0 Iteration 3 Loss function: 15589522.0 Iteration 4 Loss function: 14332520.0 Iteration 5 Loss function: 13731277.0 Iteration 6 Loss function: 13459373.0 Iteration 7 Loss function: 13346343.0 Iteration 8 Loss function: 13302241.0 Iteration 9 Loss function: 13285422.0 Iteration 10 Loss function: 13278990.0 Iteration 11 Loss function: 13276512.0 Iteration 12 Loss function: 13275553.0 Iteration 13 Loss function: 13275184.0 Iteration 14 Loss function: 13275036.0 Iteration 15 Loss function: 13274984.0 Iteration 16 Loss function: 13274959.0 Iteration 17 Loss function: 13274950.0 Iteration 18 Loss function: 13274948.0
Compute substracted signal¶
# The subtracted signal is given by Y_sub = AMBER.Ygrid - AMBER.b_grid Y_back = AMBER.b_grid # reshape the data to fit Y_sub = Y_sub.reshape(AMBER.E_size,AMBER.Qx_size,AMBER.Qy_size).T Y_back = Y_back.reshape(AMBER.E_size,AMBER.Qx_size,AMBER.Qy_size).T # reshape the observation Y_obs = AMBER.Ygrid.reshape(AMBER.E_size,AMBER.Qx_size,AMBER.Qy_size).T
Display substracted signal¶
# reshape the observation Y_obs = AMBER.Ygrid.reshape(AMBER.E_size,AMBER.Qx_size,AMBER.Qy_size).T fig0 = plt.figure(figsize=(20, 9)) # Plot 1: Observations Y ax0 = fig0.add_subplot(1, 2, 1) energy = 4.5 EIdx = np.argmin(np.abs(E-energy)) ax0.pcolormesh(H[:,:,EIdx],L[:,:,EIdx],Y_obs[:,:,EIdx],vmin=-2,vmax=50) ax0.set_xlabel('H [r.l.u.]') ax0.set_ylabel('L [r.l.u.]') ax0.set_title('Observations') # Plot 2: Subtracted Y ax1 = fig0.add_subplot(1, 2, 2) energy = 4.5 EIdx = np.argmin(np.abs(E-energy)) p = ax1.pcolormesh(H[:,:,EIdx],L[:,:,EIdx],Y_sub[:,:,EIdx],vmin=-2,vmax=50) ax1.set_xlabel('H [r.l.u.]') ax1.set_ylabel('L [r.l.u.]') ax1.set_title('Subtracted signal') fig0.colorbar(p) plt.tight_layout() plt.show()
fig0 = plt.figure(figsize=(20, 9)) # Plot 1: Observations Y ax0 = fig0.add_subplot(1, 2, 1) h_value = 0.25 HIdx = np.argmin(np.abs(h-h_value)) ax0.pcolormesh(L[:,HIdx,:],E[:,HIdx,:],Y_obs[:,HIdx,:],vmin=-2,vmax=50) ax0.set_xlabel('H [r.l.u.]') ax0.set_ylabel('L [r.l.u.]') ax0.set_title('Observations') # Plot 2: Subtracted Y ax1 = fig0.add_subplot(1, 2, 2) HIdx = np.argmin(np.abs(h-h_value)) p = ax1.pcolormesh(L[:,HIdx,:],E[:,HIdx,:],Y_sub[:,HIdx,:],vmin=-2,vmax=50) ax1.set_xlabel('L [r.l.u.]') ax1.set_ylabel('E [meV]') ax1.set_title('Subtracted signal') fig0.colorbar(p) plt.tight_layout() plt.show()
Compare the signal with the subtracted signal in a 1D cut¶
fig0 = plt.figure() # Plot 1: Observations Y ax0 = fig0.add_subplot(1, 1, 1) h_value = 0.25 e_value = 5.5 # meV EIdx = np.argmin(np.abs(e-e_value)) HIdx = np.argmin(np.abs(h-h_value)) ax0.scatter(L[:,HIdx,EIdx],Y_obs[:,HIdx,EIdx],label='Observations') ax0.scatter(L[:,HIdx,EIdx],Y_sub[:,HIdx,EIdx],label='Subtracted Signal') ax0.scatter(L[:,HIdx,EIdx],Y_back[:,HIdx,EIdx],label='Background') ax0.set_xlabel('L [r.l.u.]') ax0.set_ylabel('I [Arb. unit]') ax0.legend() plt.tight_layout() plt.show()