Sensitivity Analysis

Test here: (local) sensitivity analysis of kinematic parameters with respect to a defined objective function. Aim: test how sensitivity the resulting model is to uncertainties in kinematic parameters to:

  1. Evaluate which the most important parameters are, and to
  2. Determine which parameters could, in principle, be inverted with suitable information.

Theory: local sensitivity analysis

Basic considerations:

  • parameter vector \vec{p}
  • residual vector \vec{r}
  • calculated values at observation points \vec{z}
  • Jacobian matrix J_{ij} = \frac{\partial \vec{z}}{\partial \vec{p}}

Numerical estimation of Jacobian matrix with central difference scheme (see Finsterle):

J_{ij} = \frac{\partial z_i}{\partial p_j} \approx \frac{z_i(\vec{p}; p_j + \delta p_j) - z_i(\vec{p};p_j - \delta p_j)}{2 \delta p_j}

where \delta p_j is a small perturbation of parameter j, often as a fraction of the value.

Defining the responses

A meaningful sensitivity analysis obviously depends on the definition of a suitable response vector \vec{z}. Ideally, these responses are related to actual observations. In our case, we first want to determine how sensitive a kinematic structural geological model is with respect to uncertainties in the kinematic parameters. We therefore need calculatable measures that describe variations of the model.

As a first-order assumption, we will use a notation of a stratigraphic distance for discrete subsections of the model, for example in single voxets for the calculated model. We define distance d of a subset \omega as the (discrete) difference between the (discrete) stratigraphic value of an ideal model, \hat{s}, to the value of a model realisation s_i:

d(\omega) = \hat{s} - s_i

In the first example, we will consider only one response: the overall sum of stratigraphic distances for a model realisation r of all subsets (= voxets, in the practical sense), scaled by the number of subsets (for a subsequent comparison of model discretisations):

D_r = \frac{1}{n} \sum_{i=1}^n d(\omega_i)

Note: mistake before: not considering distances at single nodes but only the sum - this lead to “zero-difference” for simple translation! Now: consider more realistic objective function, squared distance:

r = \sqrt{\sum_i (z_{i calc} - z_{i ref})^2}

from IPython.core.display import HTML
css_file = 'pynoddy.css'
HTML(open(css_file, "r").read())
%matplotlib inline

Setting up the base model

For a first test: use simple two-fault model from paper

import sys, os
import matplotlib.pyplot as plt
import numpy as np
# adjust some settings for matplotlib
from matplotlib import rcParams
# print rcParams
rcParams['font.size'] = 15
# determine path of repository to set paths corretly below
repo_path = os.path.realpath('../..')
import pynoddy.history
import pynoddy.events
import pynoddy.output
reload(pynoddy.history)
reload(pynoddy.events)
nm = pynoddy.history.NoddyHistory()
# add stratigraphy
strati_options = {'num_layers' : 8,
                  'layer_names' : ['layer 1', 'layer 2', 'layer 3', 'layer 4', 'layer 5', 'layer 6', 'layer 7', 'layer 8'],
                  'layer_thickness' : [1500, 500, 500, 500, 500, 500, 500, 500]}
nm.add_event('stratigraphy', strati_options )

# The following options define the fault geometry:
fault_options = {'name' : 'Fault_W',
                 'pos' : (4000, 3500, 5000),
                 'dip_dir' : 90,
                 'dip' : 60,
                 'slip' : 1000}

nm.add_event('fault', fault_options)
# The following options define the fault geometry:
fault_options = {'name' : 'Fault_E',
                 'pos' : (6000, 3500, 5000),
                 'dip_dir' : 270,
                 'dip' : 60,
                 'slip' : 1000}

nm.add_event('fault', fault_options)
history = "two_faults_sensi.his"
nm.write_history(history)
output_name = "two_faults_sensi_out"
# Compute the model
pynoddy.compute_model(history, output_name)
''
# Plot output
nout = pynoddy.output.NoddyOutput(output_name)
nout.plot_section('y', layer_labels = strati_options['layer_names'][::-1],
                  colorbar = True, title="",
                  savefig = False)
png

png

Define parameter uncertainties

We will start with a sensitivity analysis for the parameters of the fault events.

H1 = pynoddy.history.NoddyHistory(history)
# get the original dip of the fault
dip_ori = H1.events[3].properties['Dip']
# dip_ori1 = H1.events[2].properties['Dip']
# add 10 degrees to dip
add_dip = -20
dip_new = dip_ori + add_dip
# dip_new1 = dip_ori1 + add_dip

# and assign back to properties dictionary:
H1.events[3].properties['Dip'] = dip_new
reload(pynoddy.output)
new_history = "sensi_test_dip_changed.his"
new_output = "sensi_test_dip_changed_out"
H1.write_history(new_history)
pynoddy.compute_model(new_history, new_output)
# load output from both models
NO1 = pynoddy.output.NoddyOutput(output_name)
NO2 = pynoddy.output.NoddyOutput(new_output)

# create basic figure layout
fig = plt.figure(figsize = (15,5))
ax1 = fig.add_subplot(121)
ax2 = fig.add_subplot(122)
NO1.plot_section('y', position=0, ax = ax1, colorbar=False, title="Dip = %.0f" % dip_ori)
NO2.plot_section('y', position=0, ax = ax2, colorbar=False, title="Dip = %.0f" % dip_new)

plt.show()
png

png

Calculate total stratigraphic distance

# def determine_strati_diff(NO1, NO2):
#     """calculate total stratigraphic distance between two models"""
#     return np.sum(NO1.block - NO2.block) / float(len(NO1.block))

def determine_strati_diff(NO1, NO2):
    """calculate total stratigraphic distance between two models"""
    return np.sqrt(np.sum((NO1.block - NO2.block)**2)) / float(len(NO1.block))



diff = determine_strati_diff(NO1, NO2)
print(diff)
5.56205897128

Function to modify parameters

Multiple event parameters can be changed directly with the function change_event_params, which takes a dictionarly of events and parameters with according changes relative to the defined parameters. Here a brief example:

# set parameter changes in dictionary

changes_fault_1 = {'Dip' : -20}
changes_fault_2 = {'Dip' : -20}
param_changes = {2 : changes_fault_1,
                 3 : changes_fault_2}
reload(pynoddy.history)
H2 = pynoddy.history.NoddyHistory(history)
H2.change_event_params(param_changes)
new_history = "param_dict_changes.his"
new_output = "param_dict_changes_out"
H2.write_history(new_history)
pynoddy.compute_model(new_history, new_output)
# load output from both models
NO1 = pynoddy.output.NoddyOutput(output_name)
NO2 = pynoddy.output.NoddyOutput(new_output)

# create basic figure layout
fig = plt.figure(figsize = (15,5))
ax1 = fig.add_subplot(121)
ax2 = fig.add_subplot(122)
NO1.plot_section('y', position=0, ax = ax1, colorbar=False, title="Original Model")
NO2.plot_section('y', position=0, ax = ax2, colorbar=False, title="Changed Model")

plt.show()
png

png

Full sensitivity analysis

Perform now a full sensitivity analysis for all defined parameters and analyse the output matrix. For a better overview, we first create a function to perform the sensitivity analysis:

import copy
new_history = "sensi_tmp.his"
new_output = "sensi_out"
def noddy_sensitivity(history_filename, param_change_vals):
    """Perform noddy sensitivity analysis for a model"""
    param_list = [] # list to store parameters for later analysis
    distances = [] # list to store calcualted distances
    # Step 1:
    # create new parameter list to change model
    for event_id, event_dict in param_change_vals.items(): # iterate over events
        for key, val in event_dict.items(): # iterate over all properties separately
            changes_list = dict()
            changes_list[event_id] = dict()
            param_list.append("event_%d_property_%s" % (event_id, key))
            for i in range(2):
                # calculate positive and negative values
                his = pynoddy.history.NoddyHistory(history_filename)
                if i == 0:
                    changes_list[event_id][key] = val
                    # set changes
                    his.change_event_params(changes_list)
                    # save and calculate model
                    his.write_history(new_history)
                    pynoddy.compute_model(new_history, new_output)
                    # open output and calculate distance
                    NO_tmp = pynoddy.output.NoddyOutput(new_output)
                    dist_pos = determine_strati_diff(NO1, NO_tmp)
                    NO_tmp.plot_section('y', position = 0, colorbar = False,
                                        title = "Dist: %.2f" % dist_pos,
                                        savefig = True,
                                        fig_filename = "event_%d_property_%s_val_%d.png" \
                                        % (event_id, key,val))
                if i == 1:
                    changes_list[event_id][key] = -val
                    his.change_event_params(changes_list)
                    # save and calculate model
                    his.write_history(new_history)
                    pynoddy.compute_model(new_history, new_output)
                    # open output and calculate distance
                    NO_tmp = pynoddy.output.NoddyOutput(new_output)
                    dist_neg = determine_strati_diff(NO1, NO_tmp)
                    NO_tmp.plot_section('y', position=0, colorbar=False,
                                        title="Dist: %.2f" % dist_neg,
                                        savefig=True,
                                        fig_filename="event_%d_property_%s_val_%d.png" \
                                        % (event_id, key,val))
            # calculate central difference
            central_diff = (dist_pos + dist_neg) / (2.)
            distances.append(central_diff)
    return param_list, distances

As a next step, we define the parameter ranges for the local sensitivity analysis (i.e. the \delta p_j from the theoretical description above):

changes_fault_1 = {'Dip' : 1.5,
                   'Dip Direction' : 10,
                   'Slip': 100.0,
                   'X': 500.0}
changes_fault_2 = {'Dip' : 1.5,
                   'Dip Direction' : 10,
                   'Slip': 100.0,
                   'X': 500.0}
param_changes = {2 : changes_fault_1,
                 3 : changes_fault_2}

And now, we perform the local sensitivity analysis:

param_list_1, distances = noddy_sensitivity(history, param_changes)

The function passes back a list of the changed parameters and the calculated distances according to this change. Let’s have a look at the results:

for p,d in zip(param_list_1, distances):
    print "%s \t\t %f" % (p, d)
event_2_property_X       2.716228
event_2_property_Dip         1.410039
event_2_property_Dip Direction       2.133553
event_2_property_Slip        1.824993
event_3_property_X       3.323528
event_3_property_Dip         1.644589
event_3_property_Dip Direction       2.606573
event_3_property_Slip        1.930455

Results of this local sensitivity analysis suggest that the model is most sensitive to the X-position of the fault, when we evaluate distances as simple stratigraphic id differences. Here just a bar plot for better visualisation (feel free to add proper labels):

d = np.array([distances])
fig = plt.figure(figsize=(5,3))
ax = fig.add_subplot(111)
ax.bar(np.arange(0.6,len(distances),1.), np.array(distances[:]))
<Container object of 8 artists>
png

png

The previous experiment showed how pynoddy can be used for simple scientific experiments. The sensitivity analysis itself is purely local. A better way would be to use (more) global sensitivity analysis, for example using the Morris or Sobol methods. These methods are implemented in the Python package SALib, and an experimental implementation of this method into pynoddy exists, as well (see further notebooks on repository, note: no guaranteed working, so far!).