Introduction to Geophysical Modelling and Inversion

James Reid

GEOPHYSICAL INVERSION FOR

MINERAL EXPLORERS
ASEG-WA, SEPTEMBER 2014

@ 2014 Mira Geoscience Ltd.

Forward modelling vs. inversion
Forward Modelling: Given a model m and predicting data d

d=F(m)
F is an operator representing the governing equations relating the model and data

Data
Model
Inversion

Geophysical inversion refers to the mathematical and statistical

techniques for recovering information on subsurface physical properties
(magnetic susceptibility, density, electrical conductivity etc) from
observed geophysical data.
What is inversion?
Inversion: Recording data d and predicting model m

-1
m=F (d)

Data
Model

F-1
What is inversion?
Forward Modelling: Given a model and predicting data

d=F(m)
Inversion: Recording data and predicting model
-1
m=F (d)
Not Possible - Ill Conditioned F

Data
Model

F-1
Iterative inversion

Calculate model Compare observed

Starting model
response using and model responses,
and acquisition
forward modelling and calculate
parameters
algorithm Objective function

Each cycle Objective function Objective function

through the large small, or maximum
inversion process no. of iterations
is called an exceeded
iteration

Alter model
parameters so Inversion process
as to reduce objective function is complete:
Output final model
 How do inversions work?
Given:
This chart summarizes the - Field observations
- Error estimates
requirements for proceeding with - Ability to forward model
inversion of geophysical data. - Prior knowledge

Discretize the Earth

Each box has important implications
for successful inversion.
Choose a suitable Design model
Ability to do forward modelling data misfit norm
calculations is assumed.

Perform inversion

Evaluate results Iterate

Interpret preferred model(s)

Models

Model Types

Single Physical
Property Value Plate in a free-space
(vacuum)
Parameterized object Plate in a half space
(susceptibility, length,
depth, orientation)
Plate in a layered
Physical property
model
varies as a function
of depth

(after: Inversion for Applied Geophysics)

Models

Model Types

Physical properties
Model is unchanging change in all 3
perpendicular to directions.
2D models profile section Generalized structure

Model objects have

limited strike length Geometry Model
Geologic unit
2.5D models boundaries adjust
location to create 3D
Concatenated 1D shapes and bodies.
models

(after: Inversion for Applied Geophysics)

Why invert data?

Helps explain complex data sets

e.g. DCIP, Gravity Gradiometry, AEM, ZTEM, DHEM
Removes topography effects

Explains the data with a model(s) of the earth:

Provides a quantitative model that can be analysed
What is the depth, geometry, volume, physical property of the model features?
More easily relates to geology - easier for interpretation
 What geologic features can be determined in the model?

Can QC the data, identify problematic data

acquisition problems

Helps separate the noise from the signal in the data

estimates the noise levels
estimates depth of penetration
 Example: Target in presence of geological noise
IP data
Data are sometimes difficult to interpret

Shallow anomalies represent

chargeable boulders in till

Subtle responses are important

Recovered chargeability
Inversion result is more easily
interpretable in terms of geology
Know The Data
In order for modelling to occur, all instrument system and survey
acquisition parameters have to be known.

In general, try to do as little as possible to the data to preserve the

information

Obviously erroneous data should be removed prior to inversion.

This includes features/anomalies in the data which are not modelled by the
forward modelling algorithm e.g., IP or SPM effects in EM data etc

RUBBISH IN = RUBBISH OUT

Inverse Modelling

Modelling Styles

• Parametric – few unknowns

e.g. TEM decay

s1 1D Conductivity model
t1
15 Data 7 unknown
dB/dt
t2 s2 model parameters

(conductivity of each layer;

t3 s3 thickness of upper three layers)
time
s4
Inverse Modelling
Modelling Styles

• Parametric – few unknowns

• Generalized – many unknowns

e.g. TEM decay

1D Conductivity model
15 Data 40 unknown
dB/dt
model parameters
1D Mesh structure predefined
but smaller than expected
structure of geology.
time
 structure inferred from the
resulting model
Inverse Modelling

Modelling Styles

• Lithology based
– VP suite (Fullagar Geophysics)
– Geomodeller (Intrepid)

• Physical Property based

– UBC-GIF codes
– Geosoft Voxi
– VP suite
Inverse Modelling
Physical Property Based Modelling

• Physical property values of many individual cells are adjusted.

• General structure is recovered

e.g. Magnetic Data

3D susceptibility model

(low value cells removed)

Inverse Modelling
Physical Property Based Modelling

• Physical property values of many individual cells are adjusted.

• General structure is recovered.

10,000+ unknown
model parameters
e.g. Magnetic Data

3D Mesh structure predefined

3D susceptibility model but smaller than expected
structure of geology.

 structure inferred from the

(low value cells removed) resulting model

RESULT IS A PHYSICAL PROPERTY MODEL

CONTAINING STRUCTURE
Inverse Modelling
Lithology Based Modelling

• Provide physical properties (single value or distribution) for each

lithology and adjust the geometry to fit the data.

RESULT IS A
GEOLOGICAL
MODEL
Selected Spectrem EM Channels (Obs - blue, Calc - red)

10^6 10^6

10^5 10^5

10^4 10^4

1000 1000

100 100

Starting Model

600 600
550 550
500 500
450 450
(courtesy Anglo American) Inverted Model

600 600
550 550
500 500
450 450
Inverse Modelling
Which Modelling Style to choose?

• Depends on the geophysical method, the survey design, and the

exploration goal. Some examples might be:

• Is the goal to define the geometry/volume?

Measure the physical properties well and choose a lithologic based
inversion (e.g. VPmg)

• Is the goal to define a thickness of cover from a few TEM soundings?

Use a parametric inversion

• Is the goal to define both physical properties and geometry?

Use a generalized inversion (e.g. UBC)

• What geologic information is available that can be integrated into the

modelling?
Acceptable models and non-uniqueness

There are infinitely many models that can explain the observed data

Why is this so?

• Because there are usually more

unknowns (model parameters)
than observed data points
(underdetermined problem)

• Some physically-based non-

uniqueness

• Real data contain noise

Acceptable models and non-uniqueness

There are infinitely many models that can explain the observed data
How to chose one of infinitely many solutions?
Narrow down the range of options using prior knowledge

Geophysical prior knowledge:

Values are positive, and/or within bounds
Physical Properties: Estimates for host rock properties
Point-location values from drill hole information

Logical prior knowledge:

Find a “simple” result - as featureless as possible.
This sacrifices resolution but prevents over-interpreting the data.

Geological prior knowledge:

Character of the model (smooth, discontinuous)
Some idea of scale length (or size) of the bodies
Structural Constraints
Model norm
The model norm is a measure of the (mathematical) “size” of a model

The inversion process is an automated decision making scheme

The model norm is a way of encoding prior information in a form suitable

for mathematical optimisation – we seek the “smallest” model

The model norm is part of the solution to nonuniqueness …

Nonuniqueness is addressed by choosing the one model (from

infinitely many) that minimizes the defined model norm
Model norms

Smooth

Minimum horizontal structure

Minimum vertical structure

Minimise difference between the model and some “reference model”

Data misfit

Measure of
DIFFERENCE
What is a good measure of misfit?

If we assume errors follow a particular distribution

then a measure of total misfit between predictions

and field data can be defined d(m):
• Predictions can be considered OK when
d(m) < tolerance
• We don’t want to fit the data too closely or we are fitting noise
• Not all measures of Data Misfit are equal
What contributes to data noise?
Natural and cultural noise sources

Accuracy and precision in data measurements

Data positioning errors

Approximations made in forward modelling

1D
2D
3D
Plates
Anisotropy
Discretization of topography
Measures of misfit
Consider the problem of fitting a straight line to the data shown below:

2.5

% Copper
1.5

0.5

0
0 10 20 30 40 50
Depth in drillhole (m)

2.5
e28
The residuals are the differences 2

between the data points and the 1.5

best-fit line at each depth % Cu 1

y = 0.0285x + 0.3333
0.5

e11
They may be positive or negative 0
0 10 20 30 40 50
Depth in drillhole (m)
Measures of misfit – L1 and L2 norms

0.65

0.6

0.55

0.5

0.45

0.4

0.35

0.3
0 2 4 6 8 10
Measures of misfit – L1 and L2 norms

0.65
y = 0.0285x + 0.333
0.6

0.55

0.5

0.45

0.4

0.35

0.3
0 2 4 6 8 10

Original data

Misfit = sum of squares of residuals (L2 norm = least-squares)

Measures of misfit – L1 and L2 norms

0.65

0.6

0.55

0.5

0.45

0.4

0.35

0.3
0 2 4 6 8 10

Original data

Misfit = sum of absolute values of residuals (L1 norm)

Measures of misfit – L1 and L2 norms

0.65
y = 0.0285x + 0.333
0.6

0.55

0.5

0.45

0.4

0.35

0.3
0 2 4 6 8 10

Add an outlying data point

Measures of misfit – L1 and L2 norms

0.65
y = 0.0298x + 0.342
0.6

0.55

0.5

0.45

0.4

0.35

0.3
0 2 4 6 8 10

One outlying data point

Misfit = sum of squares of residuals (L2 norm = least-squares)

Measures of misfit – L1 and L2 norms

0.65

0.6

0.55

0.5

0.45

0.4

L1 Less affected by outliers (noise)

0.35

0.3
0 2 4 6 8 10

One outlying data point

Misfit = sum of absolute values of residuals (L1 norm)

Combining model norms and misfit
A statement of the inverse problem is:

Find the model which

Minimises the model norm (M), and

Produces an acceptably small data misfit (D)

Mathematically, this becomes a single optimisation

“Minimise  = D + b M ” (subject to d < tolerance)

 is the combined objective function

b is the trade off parameter (regularisation parameter)
b is the regularization parameter  b
d

• Solve: *d
 (m) = d (m) + b m(m) b0

• b too large  underfitting the data. m

Structural information lost. Tikhonov curve

• b too small  overfitting the data.

Noise becomes imaged as structure.
• b just right (d  N )  optimal fit.
Best estimate of a model which
adequately re-creates the observations.
Inverse Modelling
Non-Uniqueness: Solution (partial…)

• Provide explicit geological information

• Constraints

• Combine information from independent geophysical methods

• Joint or Cooperative Inversions
e.g. Gravity with Magnetics, Airborne EM with CSAMT, etc.
Integrated Modelling: Constraints

Sources of Data
• Geologic Mapping
• DH geological logs
• Interpreted cross-sections
• 3D geological models
• Physical property data per lithology
• Located physical property data measurements

Some information is subjective and some information is objective.

As with the geophysical data we would desire to quantify the uncertainty

associated with this data as an input to the inversion.
Rock properties are the link between geology and geophysics
Shameless plug – Mira Geoscience Rock
Property Database System
http://rpds.mirageoscience.com/

Free!

6 million measurements, including GSC database and published data

Rock Property Database System

Organise, understand, preserve and provide access to physical

property data
Common Earth Modelling

Extending the model to include multiple properties,

that honour multiple data sets, on a single model
object.

Goal:
Obtain the most complete representation of the earth.

Benefits:
Improved resolution away from constraints

Allows more precise exploration using quantitative 3D

GIS analysis.
Common Earth Modelling: Constrained Inversion Modelling
2D Gravity Synthetic

(Nick Williams)
Common Earth Modelling: Constrained Inversion Modelling
2D Gravity Synthetic

(Nick Williams)
Common Earth Modelling: Constrained Inversion Modelling
2D Gravity Synthetic

(Nick Williams)
Common Earth Modelling: Constrained Inversion Modelling
2D Gravity Synthetic

Surface constraints
can result in dramatic
improvements

(Nick Williams)
Joint and cooperative inversion
Inversion using more than one geophysical
method
Methods sensitive to same physical
property (e.g., TEM and CSAMT)

Methods sensitive to related properties

(e.g., seismic and gravity)

Joint inversion – single objective function

Cooperative inversion – iterative/sequential

approach

These approaches require that we establish

relationships between the physical properties
each method is sensitive to
 Appraisal – How good is our model?
Over-fitting vs under-fitting data
Limits to the data
Limits to the physics
Depth of investigation

Suite of models

Point-spread functions
Model resolution analysis
Sensitivity analysis
Extremal models
Model Covariance Matrix
Co-Kriging error
Summary and conclusion
Inversion has the potential to greatly improve the geological
interpretation of geophysical data

• High quality data is essential for the success of geophysical modelling

• More appropriate/efficient surveys can be designed
• Complex data sets can be understood (DH IP, 3D EM)

Understanding physical property data is the key to successful

inversion interpretation.
• Rock type
• Alteration
• Mineralization
Summary and conclusion
Non uniqueness in inversion is dealt with by imposing constraints
• Provide the constraints or they will be provided for you
• Minimum structure or geological

Interpretation of inversion requires understanding of which parts

of the model are driven by constraints and which parts are driven
by data.
• Requires inspection of multiple models

Inspect observed and predicted data before accepting a model.

• Did the inversion fit the data anomalies you are interested in?
• Beware of over-fitting and under-fitting your data
 Summary and conclusion

Geologically constrained inversion will greatly improve your results

• Constraints can be factual or conceptual (hypothesis testing)

• Sparse or detailed
• From different sources

Geological maps
Outcrop samples
Estimates of overburden depth
Detailed drill data
Acknowledgements

Nigel Phillips
Dianne Mitchinson
Scott Napier - Mira Geoscience, Vancouver
Shannon Frey
Thomas Campagne

Ross Brodie - Geoscience Australia

Ken Witherly - Condor Consulting

Regis Neroni - FMGL

Doug Oldenburg - UBC-GIF

Reference

Inversion for Applied Geophysics

http://www.eos.ubc.ca/research/ubcgif/iag/iag-outline.htm