This notebook is part of the supplementary material to Computational Seismology: A Practical Introduction, Oxford University Press, 2016.
This exercise covers the following aspects:
The acoustic wave equation in 2D is $$ \ddot{p}(x,z,t) \ = \ c(x,z)^2 (\partial_x^2 p(x,z,t) + \partial_z^2 p(x,z,t)) \ + s(x,z,t) $$
and we replace the time-dependent (upper index time, lower indices space) part by
$$ \frac{p_{j,k}^{n+1} - 2 p_{j,k}^n + p_{j,k}^{n-1}}{\mathrm{d}t^2} \ = \ c_j^2 ( \partial_x^2 p + \partial_z^2 p) \ + s_{j,k}^n $$solving for $p_{j,k}^{n+1}$. The extrapolation scheme is $$ p_{j,k}^{n+1} \ = \ c_j^2 \mathrm{d}t^2 \left[ \partial_x^2 p + \partial_z^2 p \right]
Before you start it is good practice to immediately make a copy of the original notebook (e.g., X_orig.ipynb). Run the simulation code. Relate the time extrapolation loop with the numerical algorithm we developed in the course. Understand the input parameters for the simulation and the plots that are generated. Modify source and receiver locations and observe the effects on the seismograms.
Introduce a new parameter (e.g., eps) and calculate the Courant criterion. Determine numerically the stability limit of the code as accurately as possible by increasing the time step. Print the max value of the pressure field at each time step and observe the evolution of it in the case of stable and unstable simulations. (Hint: The Courant criterion is defined as $eps = (velocity * dt) / dx$ . With this information you can calculate the maximum possible, stable time step. )
Extend the code by adding the option to use a 5-point difference operator (see problem 1 of exercise sheet). Compare simulations with the 3-point and 5-point operator. Is the stability limit still the same? Make it an option to change between 3-pt and 5-pt operator. Estimate the number of points per wavelength and investigate the accuracy of the simulation by looking for signs of numerical dispersion in the resulting seismograms. The 5-pt weights are: $[-1/12, 4/3, -5/2, 4/3, -1/12]/dx^2$.
Increase the frequency of the wavefield by varying f0. Investigate the angular dependence of the wavefield. Why does the wavefield look anisotropic? Which direction is the most accurate and why? What happens if you set the source time function to a spike (zero everywhere except one element with value 1).
Now let us explore the power of the finite-difference method by varying the internal structure of the model. Here we can only modify the velocity c that can vary at each grid point (any restrictions?). Here are some suggestions. Investigate the influence of the structure by analysing the snapshots and the seismograms.
Initialize a strongly heterogeneous 2D velocity model of your choice and simulate waves propagating from an internal source point ($x_s, z_s$) to an internal receiver ($x_r, z_r$). Show that by reversing source and receiver you obtain the same seismogram.
Time reversal. Define in an arbitrary 2D velocity model a source at the centre of the domain an a receiver circle at an appropriate distance around the source. Simulate a wavefield, record it at the receiver ring and store the results. Reverse the synthetic seismograms and inject the as sources at the receiver points. What happens? Do you know examples where this principle is used?
# This is a configuration step for the exercise. Please run it before the simulation code!
%matplotlib notebook
import numpy as np
import matplotlib.pyplot as plt
Below is the 2D acoustic simulation code:
# Simple finite difference solver
# Acoustic wave equation p_tt = c^2 p_xx + src
# 2-D regular grid
nx = 200 # grid points in x
nz = 200 # grid points in z
nt = 750 # number of time steps
dx = 10.0 # grid increment in x
dt = 0.001 # Time step
c0 = 3000.0 # velocity (can be an array)
isx = nx // 2 # source index x
isz = nz // 2 # source index z
ist = 100 # shifting of source time function
f0 = 100.0 # dominant frequency of source (Hz)
isnap = 10 # snapshot frequency
T = 1.0 / f0 # dominant period
nop = 3 # length of operator
# Model type, available are "homogeneous", "fault_zone",
# "surface_low_velocity_zone", "random", "topography",
# "slab"
model_type = "homogeneous"
# Receiver locations
irx = np.array([60, 80, 100, 120, 140])
irz = np.array([5, 5, 5, 5, 5])
seis = np.zeros((len(irx), nt))
# Initialize pressure at different time steps and the second
# derivatives in each direction
p = np.zeros((nz, nx))
pold = np.zeros((nz, nx))
pnew = np.zeros((nz, nx))
pxx = np.zeros((nz, nx))
pzz = np.zeros((nz, nx))
# Initialize velocity model
c = np.zeros((nz, nx))
if model_type == "homogeneous":
c += c0
elif model_type == "fault_zone":
c += c0
c[:, nx // 2 - 5: nx // 2 + 5] *= 0.8
elif model_type == "surface_low_velocity_zone":
c += c0
c[1:10,:] *= 0.8
elif model_type == "random":
pert = 0.4
r = 2.0 * (np.random.rand(nz, nx) - 0.5) * pert
c += c0 * (1 + r)
elif model_type == "topography":
c += c0
c[0 : 10, 10 : 50] = 0
c[0 : 10, 105 : 115] = 0
c[0 : 30, 145 : 170] = 0
c[10 : 40, 20 : 40] = 0
c[0 : 15, 50 : 105] *= 0.8
elif model_type == "slab":
c += c0
c[110 : 125, 0 : 125] = 1.4 * c0
for i in range(110, 180):
c[i , i-5 : i + 15 ] = 1.4 * c0
else:
raise NotImplementedError
cmax = c.max()
# Source time function Gaussian, nt + 1 as we loose the last one by diff
src = np.empty(nt + 1)
for it in range(nt):
src[it] = np.exp(-1.0 / T ** 2 * ((it - ist) * dt) ** 2)
# Take the first derivative
src = np.diff(src) / dt
src[nt - 1] = 0
v = max([np.abs(src.min()), np.abs(src.max())])
# Initialize animated plot
image = plt.imshow(pnew, interpolation='nearest', animated=True,
vmin=-v, vmax=+v, cmap=plt.cm.RdBu)
# Plot the receivers
for x, z in zip(irx, irz):
plt.text(x, z, '+')
plt.text(isx, isz, 'o')
plt.colorbar()
plt.xlabel('ix')
plt.ylabel('iz')
plt.ion()
plt.show()
# required for seismograms
ir = np.arange(len(irx))
#################################################
# CALCULATE AND PRINT COURANT CRITERION HERE
#################################################
# Time extrapolation
for it in range(nt):
if nop==3:
# calculate partial derivatives, be careful around the boundaries
for i in range(1, nx - 1):
pzz[:, i] = (p[:, i + 1] - 2 * p[:, i] + p[:, i - 1]) / dx ** 2
for j in range(1, nz - 1):
pxx[j, :] = (p[j - 1, :] - 2 * p[j, :] + p[j + 1, :]) / dx ** 2
if nop==5:
#################################################
# IMPLEMENT 5 POINT CENTERED DIFFERENCES HERE!
#################################################
pass
# Time extrapolation
pnew = 2 * p - pold + dt ** 2 * c ** 2 * (pxx + pzz)
# Add source term at isx, isz
pnew[isz, isx] = pnew[isz, isx] + src[it]
# Plot every isnap-th iteration
if it % isnap == 0: # you can change the speed of the plot by increasing the plotting interval
plt.title("Max P: %.2f" % p.max())
image.set_data(pnew)
plt.gcf().canvas.draw()
pold, p = p, pnew
# Save seismograms
seis[ir, it] = p[irz[ir], irx[ir]]
The cell below allows you to plot source time function, seismic velocites, and the resulting seismograms in windows inside the notebook. Remember to rerun after you simulated again!
# Plot the source time function and the seismograms
plt.ioff()
plt.figure(figsize=(12, 12))
plt.subplot(221)
time = np.arange(nt) * dt
plt.plot(time, src)
plt.title('Source time function')
plt.xlabel('Time (s) ')
plt.ylabel('Source amplitude ')
plt.subplot(222)
ymax = seis.ravel().max()
for ir in range(len(seis)):
plt.plot(time, seis[ir, :] + ymax * ir)
plt.xlabel('Time (s)')
plt.ylabel('Amplitude')
plt.subplot(223)
ymax = seis.ravel().max()
for ir in range(len(seis)):
plt.plot(time, seis[ir, :] + ymax * ir)
plt.xlabel('Time (s)')
plt.ylabel('Amplitude')
plt.subplot(224)
# The velocity model is influenced by the Earth model above
plt.title('Velocity Model')
plt.imshow(c)
plt.xlabel('ix')
plt.ylabel('iz')
plt.colorbar()
plt.show()